Soft Computing Applications Proceedings of the 7th International Workshop Soft Computing Applications Sofa 2016 978-3-319-62521-8, 3319625217, 9783319625249, 3319625241, 978-3-319-62520-1

Table of contents :
Front Matter ....Pages i-xxxi
Front Matter ....Pages 1-1
Digital Sunshade Using Head-up Display (Razvan-Catalin Miclea, Ioan Silea, Florin Sandru)....Pages 3-11
Basic Expert System (Marius M. Balas, Robert A. Boran)....Pages 12-17
Expert System for Predictive Diagnosis (1) Principles and Knowledge Base (Dorin Isoc)....Pages 18-33
Expert System for Predictive Diagnosis (2) Implementing, Testing, Using (Dorin Isoc)....Pages 34-46
GA Based Multi-stage Transmission Network Expansion Planning (A. Simo, St. Kilyeni, C. Barbulescu)....Pages 47-59
Epidemic Algorithm Based Optimal Power Flow in Electric Grids (K. Muniyasamy, Seshadhri Srinivasan, S. Parthasarathy, B. Subathra, Simona Dzitac)....Pages 60-69
Issues Regarding the Tuning of a Minimum Variance Adaptive Controller (Ioan Filip, Iosif Szeidert, Octavian Prostean, Cristian Vasar)....Pages 70-77
Adaptive Controller for Networked Control Systems Subjected to Random Communication Delays (Seshadhri Srinivasan, G. Saravanakumar, B. Subathra, N. Sundarajan, Ülle Kotta, Srini Ramasamy et al.)....Pages 78-94
Aspects Regarding Risk Assessment of Human Body Exposure in Electric and Magnetic Fields (Marius Lolea, Simona Dzitac)....Pages 95-106
A Communication Viewpoints of Distributed Energy Resource (Ravish Kumar, Seshadhri Srinivasan, G. Indumathi, Simona Dzitac)....Pages 107-117
Cyber-Physical Energy Systems Approach for Engineering Power System State Estimation in Smart Grids (Seshadhri Srinivasan, Øystein Hov Holhjem, Giancarlo Marafioti, Geir Mathisen, Alessio Maffei, Giovanni Palmieri et al.)....Pages 118-132
Front Matter ....Pages 133-133
About the Applications of the Similarity of Websites Regarding HTML-Based Webpages (Doru Anastasiu Popescu, Ovidiu Domșa, Nicolae Bold)....Pages 135-142
Energy Efficient Cache Node Placement Using Genetic Algorithm with Probabilistic Delta Consistency Using Flexible Combination of Push and Pull Algorithm for Wireless Sensor Networks (Juhi R. Srivastava, T. S. B. Sudarshan)....Pages 143-163
A Fast JPEG2000 Based Crypto-Compression Algorithm: Application to the Security for Transmission of Medical Images (Med Karim Abdmouleh, Hedi Amri, Ali Khalfallah, Med Salim Bouhlel)....Pages 164-175
IoThings: A Platform for Building up the Internet of Things (Andrei Gal, Ioan Filip, Florin Dragan)....Pages 176-188
Imperialist Competition Based Clustering Algorithm to Improve the Lifetime of Wireless Sensor Network (Ali Shokouhi Rostami, Marzieh Badkoobe, Farahnaz Mohanna, Ali Asghar Rahmani Hosseinabadi, Valentina Emilia Balas)....Pages 189-202
Wireless Sensor Networks Relay Node Deployment for Oil Tanks Monitoring (Ola E. Elnaggar, Rabie A. Ramadan, Magda B. Fayek)....Pages 203-215
A Smart Way to Improve the Printing Capability of Operating System (Adeel Ahmed, Muhammad Arif, Abdul Rasheed Rizwan, Muhammad Jabbar, Zaheer Ahmed)....Pages 216-228
A Comparative Study for Ontology and Software Design Patterns (Zaheer Ahmed, Muhammad Arif, Muhammad Sami Ullah, Adeel Ahmed, Muhammad Jabbar)....Pages 229-250
Front Matter ....Pages 251-251
MainIndex Sorting Algorithm (Adeel Ahmed, Muhammad Arif, Abdul Rasheed Rizwan, Muhammad Jabbar, Zaheer Ahmed, Muhammad Sami Ullah)....Pages 253-264
Deep Learning Tools for Human Microbiome Big Data (Oana Geman, Iuliana Chiuchisan, Mihai Covasa, Cris Doloc, Mariana-Rodica Milici, Laurentiu-Dan Milici)....Pages 265-275
Evaluating the User Experience of a Web Application for Managing Electronic Health Records (Daniel-Alexandru Jurcău, Vasile Stoicu-Tivadar)....Pages 276-289
Multi Framing HDR for MRI Brain Images (Marius M. Balas, Adelin Sofrag)....Pages 290-297
Computer Aided Posture Screening Using the Microsoft Kinect on Professional Computer Users for Malicious Postures (Norbert Gal-Nadasan, Vasile Stoicu-Tivadar, Dan V. Poenaru, Diana Popa-Andrei, Emanuela Gal-Nadasan)....Pages 298-307
IT Complex Solution Supporting Continuity of Care (Mihaela Crișan-Vida, Liliana Bărbuț, Alexandra Bărbuț, Lăcrămioara Stoicu-Tivadar)....Pages 308-315
The Importance of Quantification of Data in Studies on the Health Effects of Exposure to Electromagnetic Fields Generated by Mobile Base Stations (S. M. J. Mortazavi, Valentina Emilia Balas, A. Zamani, A. Zamani, S. A. R. Mortazavi, M. Haghani et al.)....Pages 316-326
Graphical Method for Evaluating and Predicting the Human Performance in Real Time (Mariana-Rodica Milici, Oana Geman, Iuliana Chiuchisan, Laurentiu-Dan Milici)....Pages 327-338
Tremor Measurement System for Neurological Disorders Screening (Iuliana Chiuchisan, Iulian Chiuchisan, Oana Geman, Rodica-Mariana Milici, Laurentiu-Dan Milici)....Pages 339-348
Novel Method for Neurodegenerative Disorders Screening Patients Using Hurst Coefficients on EEG Delta Rhythm (Roxana Toderean (Aldea), Oana Geman, Iuliana Chiuchisan, Valentina Emilia Balas, Valeriu Beiu)....Pages 349-358
Environment Description for Blind People (J. S. Park, D. López De Luise, D. J. Hemanth, J. Pérez)....Pages 359-366
Specialized Software System for Heart Rate Variability Analysis: An Implementation of Nonlinear Graphical Methods (E. Gospodinova, M. Gospodinov, Nilanjan Dey, I. Domuschiev, Amira S. Ashour, Sanda V. Balas et al.)....Pages 367-374
Classifier Ensemble Selection Based on mRMR Algorithm and Diversity Measures: An Application of Medical Data Classification (Soraya Cheriguene, Nabiha Azizi, Nilanjan Dey, Amira S. Ashour, Corina A. Mnerie, Teodora Olariu et al.)....Pages 375-384
Personal Health Record Management System Using Hadoop Framework: An Application for Smarter Health Care (Bidyut Biman Sarkar, Swagata Paul, Barna Cornel, Noemi Rohatinovici, Nabendu Chaki)....Pages 385-393
Gene-Disease-Food Relation Extraction from Biomedical Database (Wahiba Ben Abdessalem Karaa, Monia Mannai, Nilanjan Dey, Amira S. Ashour, Iustin Olariu)....Pages 394-407
Cluster Analysis for European Neonatal Jaundice (P. K. Nizar Banu, Hala S. Own, Teodora Olariu, Iustin Olariu)....Pages 408-419
Front Matter ....Pages 421-421
Nonlinear Fourth-Order Diffusion-Based Model for Image Denoising (Tudor Barbu)....Pages 423-429
On Time-Frequency Image Processing for Change Detection Purposes (Dorel Aiordachioaie)....Pages 430-445
Adaptive Multi-round Smoothing Based on the Savitzky-Golay Filter (József Dombi, Adrienn Dineva)....Pages 446-454
Point Cloud Processing with the Combination of Fuzzy Information Measure and Wavelets (Adrienn Dineva, Annamária R. Várkonyi-Kóczy, Vincenzo Piuri, József K. Tar)....Pages 455-461
Significance of Glottal Closure Instants Detection Algorithms in Vocal Emotion Conversion (Susmitha Vekkot, Shikha Tripathi)....Pages 462-473
Analysis of Image Compression Approaches Using Wavelet Transform and Kohonen’s Network (Mourad Rahali, Habiba Loukil, Mohamed Salim Bouhlel)....Pages 474-486
Application of Higham and Savitzky-Golay Filters to Nuclear Spectra (Vansha Kher, Garvita Khajuria, Purnendu Prabhat, Meena Kohli, Vinod K. Madan)....Pages 487-496
The Efficiency of Perceptual Quality Metrics 3D Meshes Based on the Human Visual System (Nessrine Elloumi, Habiba Loukil Hadj Kacem, Med Salim Bouhlel)....Pages 497-508
Fireworks Algorithm Based Image Registration (Silviu-Ioan Bejinariu, Hariton Costin, Florin Rotaru, Ramona Luca, Cristina Diana Niţă, Camelia Lazăr)....Pages 509-523
Discrete Wavelet Transforms for PET Image Reduction/Expansion (wavREPro) (Hedi Amri, Malek Gargouri, Med Karim Abdmouleh, Ali Khalfallah, Bertrand David, Med Salim Bouhlel)....Pages 524-539
Tips on Texture Evaluation with Dual Tree Complex Wavelet Transform (Anca Ignat, Mihaela Luca)....Pages 540-556
Back Matter ....Pages 557-559

Citation preview

Advances in Intelligent Systems and Computing 633

Valentina Emilia Balas Lakhmi C. Jain Marius Mircea Balas Editors

Soft Computing Applications Proceedings of the 7th International Workshop Soft Computing Applications (SOFA 2016), Volume 1

Advances in Intelligent Systems and Computing Volume 633

Series editor Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland e-mail: [email protected]

About this Series The series “Advances in Intelligent Systems and Computing” contains publications on theory, applications, and design methods of Intelligent Systems and Intelligent Computing. Virtually all disciplines such as engineering, natural sciences, computer and information science, ICT, economics, business, e-commerce, environment, healthcare, life science are covered. The list of topics spans all the areas of modern intelligent systems and computing. The publications within “Advances in Intelligent Systems and Computing” are primarily textbooks and proceedings of important conferences, symposia and congresses. They cover significant recent developments in the field, both of a foundational and applicable character. An important characteristic feature of the series is the short publication time and world-wide distribution. This permits a rapid and broad dissemination of research results.

Advisory Board Chairman Nikhil R. Pal, Indian Statistical Institute, Kolkata, India e-mail: [email protected] Members Rafael Bello Perez, Universidad Central “Marta Abreu” de Las Villas, Santa Clara, Cuba e-mail: [email protected] Emilio S. Corchado, University of Salamanca, Salamanca, Spain e-mail: [email protected] Hani Hagras, University of Essex, Colchester, UK e-mail: [email protected] László T. Kóczy, Széchenyi István University, Győr, Hungary e-mail: [email protected] Vladik Kreinovich, University of Texas at El Paso, El Paso, USA e-mail: [email protected] Chin-Teng Lin, National Chiao Tung University, Hsinchu, Taiwan e-mail: [email protected] Jie Lu, University of Technology, Sydney, Australia e-mail: [email protected] Patricia Melin, Tijuana Institute of Technology, Tijuana, Mexico e-mail: [email protected] Nadia Nedjah, State University of Rio de Janeiro, Rio de Janeiro, Brazil e-mail: [email protected] Ngoc Thanh Nguyen, Wroclaw University of Technology, Wroclaw, Poland e-mail: [email protected] Jun Wang, The Chinese University of Hong Kong, Shatin, Hong Kong e-mail: [email protected]

More information about this series at http://www.springer.com/series/11156

Valentina Emilia Balas Lakhmi C. Jain Marius Mircea Balas •

Editors

Soft Computing Applications Proceedings of the 7th International Workshop Soft Computing Applications (SOFA 2016), Volume 1

123

Editors Valentina Emilia Balas Department of Automatics and Applied Informatics, Faculty of Engineering Aurel Vlaicu University of Arad Arad Romania

Marius Mircea Balas Department of Automatics and Applied Informatics, Faculty of Engineering Aurel Vlaicu University of Arad Arad Romania

Lakhmi C. Jain University of Canberra Canberra, ACT Australia

ISSN 2194-5357 ISSN 2194-5365 (electronic) Advances in Intelligent Systems and Computing ISBN 978-3-319-62520-1 ISBN 978-3-319-62521-8 (eBook) DOI 10.1007/978-3-319-62521-8 Library of Congress Control Number: 2017952376 © Springer International Publishing AG 2018 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Printed on acid-free paper This Springer imprint is published by Springer Nature The registered company is Springer International Publishing AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Preface

These two volumes constitute the Proceedings of the 7th International Workshop on Soft Computing Applications, or SOFA 2016, which will be held on August 24–26, 2016, in Arad, Romania. This edition was organized by Aurel Vlaicu University of Arad, Romania, University of Belgrade, Serbia, in conjunction with Institute of Computer Science, Iasi Branch of the Romanian Academy, IEEE Romanian Section, Romanian Society of Control Engineering and Technical Informatics (SRAIT)—Arad Section, General Association of Engineers in Romania—Arad Section and BTM Resources Arad. Soft computing concept was introduced by Lotfi Zadeh in 1991 and serves to highlight the emergence of computing methodologies in which the accent is on exploiting the tolerance for imprecision and uncertainty to achieve tractability, robustness, and low solution cost. Soft computing facilitates the use of fuzzy logic, neurocomputing, evolutionary computing and probabilistic computing in combination, leading to the concept of hybrid intelligent systems. Combining of such intelligent systems’ tools and a large number of applications can show the great potential of soft computing in all domains. The volumes cover a broad spectrum of soft computing techniques, theoretical and practical applications find solutions for industrial world, economic, and medical problems. The conference papers included in these proceedings, published post conference, were grouped into the following area of research: • Methods and Applications in Electrical Engineering • Knowledge-Based Technologies for Web Applications, Cloud Computing, Security Algorithms and Computer Networks • Biomedical Applications • Image, text and signal processing • Machine Learning and Applications • Business Process Management • Fuzzy Applications, Theory and Fuzzy and Control • Computational Intelligence in Education

v

vi

Preface

• Soft Computing & Fuzzy Logic in Biometrics (SCFLB) • Soft Computing Algorithms Applied in Economy, Industry and Communication Technology • Modelling and Applications in Textiles • Methods and Applications in Electrical Engineering In SOFA 2016, we had six eminent keynote speakers: Professor Michio Sugeno (Japan), Professor Anna Esposito (Italy), Professor Mika Sato-Ilic (Japan), Professor Valeriu Beiu (Romania), Professor Salil Bose (USA), Professor Rajeeb Dey (India) and an interesting roundtable of Professor József Dombi (Hungary). Their summaries talks are included in this book. We especially thank the Honorary Chair of SOFA 2016 Prof. Lotfi A. Zadeh who encouraged and motivated us, like to all the other SOFA editions. We are thankful to all the authors that have submitted papers for keeping the quality of the SOFA 2016 conference at high levels. The editors of this book would like to acknowledge all the authors for their contributions and also the reviewers. We have received an invaluable help from the members of the International Program Committee and the chairs responsible for different aspects of the workshop. We appreciate also the role of special sessions organizers. Thanks to all of them we had been able to collect many papers of interesting topics, and during the workshop, we had very interesting presentations and stimulating discussions. For their help with organizational issues of all SOFA editions we express our thanks to TRIVENT Company, Mónika Jetzin and Teodora Artimon for having customized the software Conference Manager, registration of conference participants and all local arrangements. Our special thanks go to Janus Kacprzyk (Editor in Chief, Springer, Advances in Intelligent Systems and Computing Series) for the opportunity to organize this guest edited volume. We are grateful to Springer, especially to Dr. Thomas Ditzinger (Senior Editor, Applied Sciences & Engineering Springer-Verlag) for the excellent collaboration, patience, and help during the evolvement of this volume. We hope that the volumes will provide useful information to professors, researchers, and graduated students in the area of soft computing techniques and applications, and all will find this collection of papers inspiring, informative, and useful. We also hope to see you at a future SOFA event. Valentina Emilia Balas Lakhmi C. Jain Marius Mircea Balas

Keynote Presentations

Introduction to Choquet Calculus Michio Sugeno Tokyo Institute of Technology, Japan Abstract. In this talk, we give a brief introduction to a recent topic “Choquet calculus” where calculus consists of integrals and derivatives. Modern integrals, namely Lebesgue integrals initiated by Lebesgue in 1902, are associated with the concept of “measures.” Lebesgue measures are defined as additive set functions with certain conditions, and hence, Lebesgue integrals hold additivity by inheriting the property of measures. In the latter half of the twentieth century, a new concept of “measures without additivity” named fuzzy measures was proposed by Sugeno in 1974. Fuzzy measures (non-additive measures in general) are defined as monotone set functions and considered as a natural extension of Lebesgue measures leading to the concept of non-additive integrals with respect to non-additive measures, where we note that monotonicity contains additivity in it. In 1953, Choquet studied the so-called Choquet functionals based on capacities, where capacities representing “potential energy” of a set were also monotone set functions but not captured as “measures” as in the sense of Lebesgue. Together with the appearance of fuzzy measures, Choquet functionals were finally re-formulated as non-additive integrals with respect to fuzzy measures by Höle in 1982. Since then, various non-additive integrals with respect to non-additive measures have been suggested. Among them, we focus on Choquet integrals which are the most general extension of Lebesgue integrals. Once we obtain the concept of integrals, we become curious about their inverse operations. In the case of Lebesgue integrals, Radon and Nikodym gave Radon-Nikodym derivatives as inverse operations in 1913 and 1930, respectively. It is well-known that with the aid of Radon-Nikodym derivatives, Kolmogorov proved the existence of conditional probabilities in 1933 and thus initiated modern probability theory, where probabilities are nothing but Lebesgue measures. On the other hand in fuzzy measure theory, conditional fuzzy measures have been not well defined in the sense of Kolmogorov. Very recently, inverse operations of Choquet integrals were studied as “derivatives with respect to fuzzy measures” (Sugeno 2013). In this talk, we deal with Choquet calculus (Sugeno 2015) based on Choquet integrals and derivatives. So far, most studies on Choquet integrals have been devoted to a discrete case. In Choquet calculus, we deal with continuous Choquet integrals and derivatives as well. First, we show how to calculate continuous Choquet integrals. To this aim, we consider distorted Lebesgue measures as a special class of fuzzy measures, and nonnegative and non-decreasing functions. The Laplace transformation is used as a basic tool for calculations. Distorted Lebesgue measures are obtained by monotone transformations of Lebesgue measures according to the idea of distorted probabilities suggested by Edwards in 1953. We remember that Kahneman was awarded Nobel Prize in Economics in 2002, where his cumulative prospect theory is based on “Choquet integrals with respect to distorted probabilities.” Next, we define

ix

x

M. Sugeno

derivatives of functions with respect to distorted Lebesgue measures, where the derivatives correspond to Radon-Nikodym derivatives in the case of Lebesgue integrals. We also discuss the identification of distorted Lebesgue measures which is a problem arising particularly in Choquet calculus. Then, we show some relations between Choquet calculus and fractional calculus which is recently getting very popular, for example, in control theory. Also, we consider differential equations with respect to distorted Lebesgue measures and give their solutions. Lastly, we present the concept of conditional distorted Lebesgue measures defined with the aid of Radon-Nikodym-like derivatives.

Biography Michio Sugeno was born in Yokohama, Japan, in 1940. After graduating from the Department of Physics, the University of Tokyo, he worked at Mitsubishi Atomic Power Industry. Then, he served the Tokyo Institute of Technology as research associate, associate professor, and professor from 1965 to 2000. After retiring from the Tokyo Institute of Technology, he worked as Laboratory Head at the Brain Science Institute, RIKEN, from 2000 to 2005, and Distinguished Visiting Professor at Doshisha University from 2005 to 2010 and then Emeritus Researcher at the European Centre for Soft Computing, Spain, from 2010 to 2015. He is Emeritus Professor at the Tokyo Institute of Technology. He was President of the Japan Society for Fuzzy Theory and Systems from 1991 to 1993, and also President of the International Fuzzy Systems Association from 1997 to 1999. He is the first recipient of the IEEE Pioneer Award in Fuzzy Systems with Zadeh in 2000. He also received the 2010 IEEE Frank Rosenblatt Award and Kampét de Feriét Award in 2012. His research interests are Choquet calculus, fuzzy measure theory, nonlinear control, and preference theory, applications of systemic functional linguistics and language functions of the brain.

Emotional Facial Expressions: Communicative Displays or Psychological Universals? Anna Esposito Department of Psychology, and IIASS, Seconda Università di Napoli, Italy [email protected]; [email protected] http://www.iiassvietri.it/anna.html Abstract. Emotional feelings permeate our everyday experience, consciously or unconsciously, driving our daily activities and constraining our perception, actions, and reactions. During daily interactions, our ability to decode emotional expressions plays a vital role in creating social linkages, producing cultural exchanges, influencing relationships, and communicating meanings. Emotional information is transmitted through verbal (the semantic content of a message) and nonverbal (facial, vocal, gestural expressions, and more) communicative tools and relations and exchanges are highly affected by the way this information is coded/decoded by/from the addresser/addressee. The accuracy above the chance in decoding facial emotional expressions suggested they can play the role of psychological universals. However, this idea is debated by data suggesting that they play the role of social messages dependent upon context and personal motives. These open questions are discussed in this talk, at the light of experimental data obtained from several experiments aimed to assess the role of context on the decoding of emotional facial expressions. The reported data support more the idea that facial expressions of emotions are learned to efficiently and effectively express intentions and negotiate relations, even though particular emotional aspects show similarities across cultural boundaries. Research devoted to the understanding of the perceptual and cognitive processes involved in the decoding of emotional states during interactional exchanges is particularly relevant in the field of Human-Human, Human-Computer Interaction and Robotics, for building and hardenind human relationships, and developing friendly, emotionally, and socially believable assistive technologies.

Biography Anna Esposito received her “Laurea Degree” summa cum laude in Information Technology and Computer Science from the Università di Salerno in 1989 with a thesis on: The Behavior and Learning of a Deterministic Neural Net (published on Complex System, vol 6(6), 507–517, 992). She received her PhD Degree in Applied Mathematics and Computer Science from the Università di Napoli,

xi

xii

A. Esposito

“Federico II” in 1995. Her PhD thesis was on: Vowel Height and Consonantal Voicing Effects: Data from Italian (published on Phonetica, vol 59(4), 197–231, 2002) and was developed at Massachusetts Institute of Technology (MIT), Research Laboratory of Electronics (RLE), under the supervision of Professor Kenneth N Stevens. She has been a Postdoc at the International Institute for Advanced Scientific Studies (IIASS), and Assistant Professor at the Department of Physics at the Università di Salerno (Italy), where she taught courses on cybernetics, neural networks, and speech processing (1996– 2000). She had a position as Research Professor (2000–2002) at the Department of Computer Science and Engineering at Wright State University (WSU), Dayton, OH, USA. She is currently associated with WSU as Research Affiliate. Anna is currently working as an Associate Professor in Computer Science at the Department of Psychology, Seconda Università di Napoli (SUN). Her teaching responsibilities include cognitive and algorithmic issues of multimodal communication, human–machine interaction, cognitive economy, and decision making. She authored 160+ peer-reviewed publications in international journals, books, and conference proceedings. She edited/co-edited 21 books and conference proceedings with Italian, EU, and overseas colleagues. Anna has been the Italian Management Committee Member of: • COST 277: Nonlinear Speech Processing, http://www.cost.esf.org/domains_ actions/ict/Actions/277 (2001–2005) • COST MUMIA: Multilingual and Multifaceted Interactive information Access, • www.cost.esf.org/domains_actions/ict/Actions/IC1002 (2010–2014) • COST TIMELY: Time in Mental Activity, www.timely-cost.eu (2010–2014) • She has been the proposer and chair of COST 2102: Cross Modal Analysis of Verbal and Nonverbal Communication, http://www.cost.esf.org/domains_ actions/ict/Actions/2102 (2006–2010). Since 2006, she is a Member of the European Network for the Advancement of Artificial Cognitive Systems, Interaction and Robotics (www.eucognition.org); She is currently the Italian Management Committee Member of ISCH COST Action IS1406: Enhancing children’s oral language skills across Europe and beyond (http:// www.cost.eu/COST_Actions/isch/Actions/IS1406) Anna’s research activities are on the following three principal lines of investigations: • 1998 to date: Cross-modal analysis of speech, gesture, facial and vocal expressions of emotions. Timing perception in language tasks. • 1995 to date: Emotional and social believable Human-Computer Interaction (HCI). • 1989 to date: Neural Networks: learning algorithm, models and applications.

Soft Data Analysis Based on Cluster Scaling Mika Sato-Ilic University of Tsukuba, Japan [email protected] Abstract. There is a growing need to analyze today’s vast and complex societal data; however, conventional data analysis that is dependent on statistical methods cannot deal with the frequently complex data types that make up this data. As early as 2000, the following six challenges were reported as important future challenges in the core area statistical research in the twenty-first century. The six challenges pointed out are (1) scales of data, (2) data reduction and compression, (3) machine learning and neural networks, (4) multivariate analysis for large p, small n (high dimension low sample size data), (5) Bayes and biased estimation, and (6) middle ground between proof and computational experiment. Soft data analysis which is soft computing-based multivariate analysis is the core area in which to combine conventional statistical methods and machine learning or data mining methods, and has a strong capability to solve the statistical challenges in the twenty-first century. In soft data analysis, we have developed cluster-scaled models which use the obtained cluster as the latent scale for explaining data. While the original scale does not have the capacity to work as the scale for complex data, a scale that is extracted from the data itself will have the capability to deal with the vast and complex data. This presentation outlines the problems and challenging issues of statistical data analysis caused by the new vast and complex data, how our cluster-scaled models are related with these issues, and how the models solve the problems with some applications.

Biography Prof. Mika Sato-Ilic currently holds the position of Professor in the Faculty of Engineering, Information and Systems, at the University of Tsukuba, Japan. She is the founding Editor in Chief of the International Journal of Knowledge Engineering and Soft Data Paradigms, Associate Editor of Neurocomputing, Associate Editor of Information Sciences, Regional Editor of International Journal on Intelligent Decision Technologies, and Associate Editor of the International Journal of Innovative Computing, Information and Control Express Letters, as well asserving on the editorial board of several other journals.In addition, she was a Council of the International Association for xiii

xiv

M. Sato-Ilic

Statistical Computing (a Section of the International Statistical Institute), a Senior Member of the IEEE, where she held several positions including the Vice-Chair of the Fuzzy Systems Technical Committee of the IEEE Computational Intelligence Society. In addition, she has served on several IEEE committees including the administration committee, program co-chair, and special sessions co-chair. Her academic output includes four books, 10 chapters, and over 120 journal and conference papers. Her research interests include the development of methods for data mining, multidimensional data analysis, multimode multiway data theory, pattern classification, and computational intelligence techniques for which she has received several academic awards.

Why the Brain Can and the Computer Can’t Biology Versus Silicon Valeriu Beiu Universitatea “Aurel Vlaicu” din Arad, Romania [email protected] Abstract. This presentation aims to follow on von Neumann’s prescient “The Computer and the Brain” (Yale Univ. Press, 1958) and—relying on the latest discoveries—to explain how the Brain achieves ultra-low power and outstandingly high reliability (nano-)computing, while our silicon-based computers cannot. The tale will be reversed as starting from the brain, and in particular from very fresh experimental results, for information processing and communication. We shall proceed from the gated ion channels which are the nano-switches of the brain. Understanding the ways, ion channels communicate will allow analyzing the statistical behaviors of arrays of gated ion channels. These will be followed by unexpected results showing that highly reliable communication using arrays of absolutely random devices is possible at amazingly small redundancy factors ( > < y ¼ y0 þ mt ;t 2 R z ¼ z0 þ nt > > : Ox þ My þ Nz þ P ¼ 0 If Ol þ Mm þ Nn 6¼ 0 means that the system has an unique solution, in other words the intersection between the line and the plan is a single point. B. Eye Tracking System After the exterior detection was performed the system has to do an interior detection, to find out the clear position of the driver. The eyes and the face of the driver have to be protected against glare, that’s why we decided to use an eye tracking system for this evaluation. An eye tracking system is usually used to monitor the driver’s loss of attention. It starts with a face detection followed by eye detection and the last step is the evaluation of the eye state [7]. If the face is turned side or the eyes are closed for a few seconds the system alerts the driver by flashing some interior light or through some vibrations in the steering wheel [12]. The head’s and eyes’ movements are monitored with cameras usually equipped with active infrared illumination. The cost of an eye tracking system is high at this moment. Such systems are not installed on the commercial vehicles only the high-end ones dispose of them. In [8] Hong et al. proposed a low cost eye tracking system which has to be placed on the driver’s head. The gaze of the subject is monitored using two cameras and two mirrors installed on the side of the head. Our system needs only the information regarding the position of the head and eyes in order to protect them from the undesired glare phenomenon. So, if the vehicle is not equipped with an eye tracking system an after-market one can be used to point out the position of the driver, like the one presented above. C. Head-up Display Head-up Displays (HUD) are becoming more and more popular, especially on high-class vehicles. If the first variants of HUD were able to display only some basic

Digital Sunshade Using Head-up Display

9

information like speed and navigation icons, the latest variants are more complex and can exhibit 3D images, the so called augmented reality technology, such as navigation information, lane guidance or potential dangers. As it was already mentioned, head-up display were basically designed to display useful information in the driver’s field of view in order to discard any kind of distraction. A HUD with the characteristics of projecting unlimited combination of images without any fixed dial’s position [10] is the desired solution for our system in order to reduce the glare effect on driver’s vision. Comparing to most of the existing commercial head-up displays, the one used in our system has to able to deal with 3 constrains: • to be dynamically adjustable on both axis, x and y • to dynamically adjust the image opacity based on the sun brightness intensity • to dynamically adjust the size of the projected image All these three conditions can be fulfilled with the new generation of HUD, based on DLP technology [11] without any hardware change, only a few software improvements are needed which mean very low costs. In [9] Wientapper et al. proposed a head-up display variant which has as input information the viewer’s head position in order to display the image in the ideal position for the driver. If such a HUD with internal calibration becomes reality on the commercial vehicles, the complexity of our proposed system decreases, being no need of an additional eye tracking system. In Fig. 6 is illustrated the result of rejecting glare from the image. The system gathers the data from the light sensor and eye tracking system and projects an adjusted image (in opacity and size) on the windshield.

Fig. 6. Rejecting the glare phenomenon using HUD

4 System Deployment - Functionality In this section will be explained the functionality of the Digital Sunshade based the below flow chart (Fig. 7). Our proposal is to let the driver to decide if he wants to use the head-up display functionality or the digital sunshade functionality because both functionalities can’t work in parallel.

10

R.-C. Miclea et al.

Fig. 7. Digital sunshade’s functionality flow chart

When the digital sunshade functionality is chosen, the sun tracking system starts the monitoring. If on the windshield’s surface is detected brighter light than the predefined threshold, the system gets the data from the sun sensor regarding the sun position. Beside the information regarding to the sun position, the sun tracking system has to offer also information regarding the light intensity level and the size of the spot which has to be covered. When the system detects light on the windshield it automatically starts also an interior detection, to find out if the sun rays are disturbing the driver. If they do, all these information are sent to the ECU in order to initiate the projection process. The head-up display integrated in this system has to be able to project light in every point of the windshield. The ECU configures the image which has to be displayed based on the information gathered from the sun tracking system (position, brightness level and size of the spot) and eye tracking system (driver’s position). These data are sent to the head-up display which projects an adapted image (opacity and size) so that the entire disturbing area to be covered (see Fig. 6).

5 Conclusions Sometimes light and solar energy is useful, such as for agriculture [13], but the car driving sunshine definitely bothers us. In this paper is presented a new concept of an anti-glare system able to protect the driver against the annoying sun rays no matter where these are crossing the windshield. Nowadays the solution for the glare problem is the sun visor, but its biggest drawback is that it can protect only the top area of the windshield. Unfortunately the glare phenomenon is unpredictable; we have to go against the sun rays crossing directly the windshield surface but also with sun rays reflected from a wet or icy road. For the latter category the sun visor is useless. The automotive companies are working hard to find a solution for this problem; the one proposed until now, with an integrated thin film, is expensive from the manufacturing point of view.

Digital Sunshade Using Head-up Display

11

We analyzed the problem together with the state of the art and we concluded that the proposed method can be a cost-efficient variant. The system is able to protect the driver against glare, no matter where the rays are hitting the windshield, and is also reasonable from the cost point of view because we used equipments already installed in the vehicle, we only gave them new functionalities: using the GPS system and some light sensors we detected the sun position, with the eye tracking system we indicated the driver’s position and in the last phase with the help of a head-up display we project an image on the windshield in the disturbing area. Furthermore, using the same principles exhibited above, the system can be improved such as to become useful also for headlamps glaring.

References 1. Viereck, V., Ackerman, J., Li, Q., Jaker, A., Schmid, J., Hillmer, H.: Sun glasses for buildings based on micro mirror arrays: technology, controled by networked sensors and scaling potential. In: 5th International Conference on Network Sensing Systems, INSS 2008, Kanazawa (2008) 2. Elahi, A.H.M.F., Rahman, M.S.: Intelligent windshield for automotive vehicles. In: 17th International Conference on Computer and Information Technology (ICCIT), Dhaka (2014) 3. Pomerleau, D.A.: Visibility estimation from a moving vehicle using the RALPH vision system. In: IEEE Conference on Intelligent Transportation Systems, ITCS 1997, Boston (1997) 4. Wu, S., Wen, C., Luo, H., Chen, Y., Wang, C., Li, J.: Using mobile lidar point clouds for traffic sign detection and sign visibility estimation. In: 2015 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Milan (2015) 5. Sato, R., Doman, K., Deguchi, D., Mekada, Y.: Visibility estimation on traffic signals under rainy weather conditions for smart driving support. In: 2012 15th International IEEE Conference on Intelligent Transportation Systems (ITSC), Anchorage (2012) 6. Hauliere, N., Tarel, J.P., Lavenant, J., Aubert, D.: Automatic fog detection and estimation of visibility distance through use of an onboard camera. Mach. Vis. Appl. 17, 8–20 (2006) 7. Dasgupta, A., George, A., Happy, S.L., Routray, A.: A vision-based system for monitoring the loss of attention in automotive drivers. IEEE Trans. Intell. Transp. Syst. 14, 1825–1838 (2013) 8. Hong, A.K.A., Pelz, J., Cockburn, J.: Lightweight, low cost, side-mounted mobile eye tracking system. In: 2012 Western New York Image Processing Workshop (WNYIPW), (2012) 9. Wientapper, F., Wuest, H., Rojtberg, P., Fellner, D.: A camera-based calibration for automotive augmented reality head-up-displays. In: 2013 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), Adelaide (2013) 10. Charissis, V., Naef, M.: Evaluation of prototype automotive head-up display interface: testing driver’s focusing ability through a VR simulation. In: 2007 IEEE Intelligent Vehicles Symposium, Istanbul (2007) 11. Texas Instruments. Enabling the next generation of Automotive Head-up Display Systems (2013) 12. Kang, H.-B.: Monitoring driver’s state and predicting unsafe driving behavior. In: Algorithm & SoC Design for Automotive Vision Systems, pp. 143–161. Springer Science + Business Media, Dordrecht (2014) 13. Nanu, S., Sumalan, R.L.: Solar Irrigation System, Patent Number: RO130028-A2, Patent Assignee: Rosenc Clusterul Energii Sustenabile, 27 Feb 2015

Basic Expert System Marius M. Balas(&) and Robert A. Boran “Aurel Vlaicu” University of Arad, Arad, Romania [email protected], [email protected]

Abstract. The paper is presenting a new visual software tool for developing expert systems: Basic Expert System. The Graphical User Interface is organized in expert diagrams, facilitating and accelerating the design of complex applications. An expert system for tomato disease diagnostic illustrates the BES operation. Keywords: Expert system Tomato disease




Visual programming language




Dataflow




1 Introduction We are living in an era of computers and machines and they all communicate with each other using binary information “0” and “1”, very fast and easy for them, but inappropriate and hard to understand for humans. We have been using computers power and speed to solve a lot of our problems but more and more we realize that there still exist problems that somehow can be better solved by humans. Instead of numerical algorithms we humans use intuition and world knowledge to solve our problems and this is very different from computer’s way of thinking. Therefore we have a gap between our human way of doing things and the computers way. A classical solution to fill this gap is represented by the visual programming languages (VPL) that lets users create their applications graphically, by manipulating basic program elements. In VPLs textual instructions are replaced by spatial arrangements of graphic symbols (blocks) allows programming with visual expressions. The simplest VPL (known as dataflow) are based on “boxes and arrows” approach, where entities are represented as boxes and the relations between them by arrows, lines or arcs. The objective of our work is to facilitate the design and the development of general purpose expert systems with a very friendly, fast and portable VPL.

2 On the Expert Systems Expert systems were introduced by the Stanford Heuristic Programming Project led by Edward Feigenbaum, who is sometimes referred as the “father of expert systems”. E. Feigenbaum, said that the key insight of early expert systems was that “intelligent

© Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_2

Basic Expert System

13

systems derive their power from the knowledge they possess rather than from the specific formalisms and inference schemes they use” [1]. The Stanford researchers tried to identify domains where expertise was highly valued and complex, such as diagnosing infectious diseases (Mycin) and identifying unknown organic. Until then, research had been focused on attempts to develop very general-purpose problem solvers such as those described by Allen Newell and Herb Simon [2]. In addition to Feigenbaum key early contributors were Edward Shortliffe, Bruce Buchanan, and Randall Davis. Expert systems were among the first truly successful forms of AI software [3–7]. Research on expert systems was also active in France. In the US the focus tended to be on rule-based systems, first on systems hard coded on top of LISP programming environments and then on expert system shells developed by vendors such as IntelliCorp. In France research focused more on systems developed in Prolog. The advantage of expert system shells was that they were easier for non-programmers to use. The advantage of Prolog environments was that they weren’t focused only on IF-THEN rules. Prolog environments provided a much fuller realization of a complete First Order Logic environment [x]. In the 1980s, expert systems proliferated. Universities offered expert system courses and two thirds of the Fortune 1000 companies applied the technology in daily business activities. Interest was international with the Fifth Generation Computer Systems project in Japan and increased research funding in Europe. In 1981 the first IBM PC was introduced, with the MS-DOS operating system. The imbalance between the relatively powerful chips in the highly affordable PC compared to the much more expensive price of processing power in the mainframes that dominated the corporate IT world at the time created a whole new type of architecture for corporate computing known as the client-server model. Using a PC, calculations and reasoning could be performed at a fraction of the price of a mainframe. This model also enabled business units to bypass corporate IT departments and directly build their own applications. As a result, client server had a tremendous impact on the expert systems market. Expert systems were already outliers in much of the business world, requiring new skills that many IT departments did not have and were not eager to develop. They were a natural fit for new PC-based shells that promised to put application development into the hands of end users and experts. Up until that point the primary development environment for expert systems had been high end Lisp machines from Xerox, Symbolics and Texas Instruments. With the rise of the PC and client server computing vendors such as IntelliCorp and Inference Corporation shifted their priorities to developing PC based tools. In addition new vendors often financed by Venture Capital started appearing regularly. These new vendors included Aion Corporation, Neuron Data, Exsys, and many others. In the 1990s and beyond the term “expert system” and the idea of a standalone AI system mostly dropped from the IT lexicon. There are two interpretations of this. One is that “expert systems failed”: the IT world moved on because expert systems didn’t

14

M.M. Balas and R.A. Boran

deliver on their over hyped promise, the fall of expert systems was so spectacular that even AI legend Rishi Sharma admitted to cheat in his college project regarding expert systems, because he didn’t consider the project worthwhile [x], the other is the mirror opposite, that expert systems were simply victims of their success. As IT professionals grasped concepts such as rule engines such tools migrated from standalone tools for the development of special purpose “expert” systems to one more tool that an IT professional has at their disposal. Many of the leading major business application suite vendors such as SAP, Siebel and Oracle integrated expert system capabilities into their suite of products as a way of specifying business logic. Rule engines are no longer simply for defining the rules an expert would use but for any type of complex, volatile, and critical business logic. They often go hand in hand with business process automation and integration environments.

3 BES, the Basic Expert System BES (Basic Expert System) is a new visual fuzzy-expert system development shell belonging to the VPL family that fills the gap between human expertise and the computer way of “thinking” by using a graphical approach instead of algorithms and code. If using BES it is possible for the human expert to create and design his own expert system by representing the knowledge base in a graphical manner, that is easy to understand and easy to make. After creating an expert system, BES runs it in real time, so the human expert may test very easy the entire application, without concerning about the inference engine that is behind the scene. A reference product when creating BES was VisiRule, created by LPA (Logic Programming Associates), and our concrete objective was to make portable standalone applications extremely fast and easy to build.

4 The BES Tomato Disorder Diagnostic Expert System Several applications were already developed with BES, which proved to be extremely friendly and effective in all senses. The most complex of these application is a Tomato Disorder Diagnostic Expert System, which is relying on the expert knowledge provided by Texas AgriLife [8] (Figs. 1, 2 and 3).

Basic Expert System

Fig. 1. The Tomato Disorder Diagnostic Expert System (overall block view)

15

16

M.M. Balas and R.A. Boran

Fig. 2. The Tomato Disorder Diagnostic Expert System (block properties setting)

Fig. 3. A diagnosis with Tomato Disorder Diagnostic Expert System

5 Discussion The paper is presenting a new visual programming language dedicated to the development of fuzzy-expert systems, Basic Expert System, conceived and successfully tested at Aurel Vlaicu University of Arad.

Basic Expert System

17

Several applications were developed in very short time (2–3 months). The implementation of an expert system for the diagnosis of tomato diseases is illustrating the operation of this new software tool. Working with BES proved to be extremely easy, fast and reliable, and our intention is to continue to develop and to refine it. Acknowledgement. This work was co-funded by European Union through European Regional Development Funds Structural Operational Program “Increasing of Economic Competitiveness” Priority axis 2, operation 2.1.2. Contract Number 621/2014.

References 1. University of Zurich. http://www.ifi.unizh.ch/groups/ailab/people/bongard/migros/LectMon830. Pdf. Accessed 12 July 2016 2. Poole, D., Macworth, A., Goebel, R.: Computational Intelligence a Logical Approach. Oxford University Press, Oxford (1998) 3. http://www.scism.sbu.ac.uk/*darlink 4. Cristea, D.: Programarea Bazată pe Reguli. Editura Academiei Romane, Bucureşti (2002) 5. Sambotin, C.: Sisteme Expert cu Prolog. Editura didactica si pedagogica, Bucureşti (1997) 6. Balas, V.E., Koprinkova-Hristova, P., Jain, L.C. (eds.): Innovations in Intelligent Machines-5. Studies in Computational Intelligence, Springer (2014) 7. Shalfield, R.: VisiRule User Guide. Logic Programming Associates, London (2008) 8. Texas AgriLife Extension Service: Tomato Problem Solver. A Guide to the Identification of Common Problems. http://aggie-horticulture.tamu.edu/vegetable/problem-solvers/tomatoproblem-solver/. Accessed 12 July 2016

Expert System for Predictive Diagnosis (1) Principles and Knowledge Base Dorin Isoc(B) Technical University of Cluj-Napoca, Cluj-Napoca, Romania [email protected]

Abstract. Implementing an expert system is a work that involves a lot of theoretical knowledge and a large amount of expertise. A particular technical application is the predictive diagnosis of underground functioning power cables. Particular are the cables modeling, but also the manner to estimate the problem of life estimation in given conditions. The work achieves the underground cable modeling by putting together determinable information. The template of knowledge piece is next built and the appropriate inference algorithm will be defined. Finally, a functional scheme of dedicated expert system is given. The conclusions emphasize that the standard structure of expert system is retrieved and significant is the project associated to knowledge. Keywords: Expert system · Underground power cable · Knowledge base · Knowledge piece · Approximate reasoning · Inference engine

1

Introduction

The correct management of electrical power asks to establish and know in each moment the state of power cables. The operation of electrical power network assumes state parameters so different that detecting of faults is practical impossible without removing the cables set out of voltage. More, knowing the behavior in time of power cables is a combined result where found it also the work regime of given subsystem. An overview of the research and reported technical solutions allows the identification of the stages already presented. It is to emphasize that a special attention is paid to technical solutions which allow the cables fault detection [1,5,7–9,11]. It is obvious that they are studied particular faults which can be detected in particular situations. In all these situations, the power cable is seen as a part of an electric circuit. By means of some ad-hoc scheme sit is possible to detect and locate the faults. The faults in power cable are short-circuits, damaged insulation, interruption. The theoretical grounds are very simple, but the technical solutions are depending of the quality of technical means, especially the sensitivity of measurements means. c Springer International Publishing AG 2018  V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8 3

Expert System for Predictive Diagnosis (1) Principles and Knowledge Base

19

So, the research goal is to develop a predictive diagnosis system. The searched technical solutions will be able that, together with the measurement means, to ensure the decisional support necessary to process data, information and knowledge able to describe the technical system of power cables. The prediction issue is an important one because once the prediction accurate achieved [2,6,10] it allows arguing decision-making processes which can be well implemented inclusively in automatic work regimes. The prediction, as anticipating technique, is less dependent of phenomena and application fields as long as the interest variables are efficiently measured. The found solutions are alternative manners to be used in different applications. The work aims to present the designing and building of an expert system dedicated to diagnose and predict the state of electrical cables in use. First one defines the problem to be solved. They are approached some theoretical grounds regarding the modeling of dynamical behavior of power cables, of the predictability issue, with direct application on hybrid models and methods able to keep the dynamics inside of hybrid models. A further chapter insist son the ways to operate with hybrid models and next section is dedicated to the functional model of power cable, model which proves to be a hybrid one. The last section includes conclusions on the manner to build and use an intelligent diagnoses and prediction system dedicated to power cables. Some points of view to be used further in order to capitalize the suggested solution are also introduced.

2 2.1

The Issue of Underground Power Cables Monitoring Sustainable Development and Technical Achievements

Sustainable development is a concept that is opposed to growth, seemingly bound-less consumerism, to the contempt for the earth’s resources. Sustainable development is a reaction to more frequent and alarming signals of those who invoke the need to protect the environment. When talking about sustainable development, a special concerned class of technical achievements are those that are built for permanent use, for a considerable time period, have a particularly important role in daily life, undertaken with significant financial and material efforts. Such achievements are the dams, the bridges, and infrastructure parts of water supply systems, power, oil and gas networks, and communication systems. All these achievements are parts of complex systems, with large size and density, with high resources concentration, often in areas of large human agglomerations. These technical systems are very similar and have some common features: (i) They require very important investments. (ii) They are serving large human communities and are unique, i.e. not necessarily involve alternative to take their services if special circumstances occur (technical functional redundancy).

20

D. Isoc

(iii) They cover services which do not involve permanent interruptions allowing lasting interventions. (iv) They are difficult to follow as operating state due to environments where they are implanted. (v) In terms of constructive manner point of view, they have special features that are rarely found in other systems, making them unique or little studied. Among these means of technical support of the civilization, the subject of current research is focusing on the underground power cables. It appears easily that these technical systems fulfill abundantly the conditions listed above. 2.2

Underground Power Cable as Technical System

Underground power cable appears as a means to transport electric energy. Of the physical point of view, the power cable acts as a resistance, at most as a electrical impedance. From the technical point of view, however, the power cable is a mechanical assembly that operates under the electric and electromagnetic fields action. The cable contains layers of different materials for roles of mechanical protection of the conductor itself as in Fig. 1. 1 2 3 4

5

6

Fig. 1. Structure of power cable in cross section (see [1]): 1 - external protective polymeric jacket; 2 - metallic shield usually made from copper foil or wire texture (screen); 3 - external semi-conducting layer; 4 - cable insulator usually made from polyethylene; 5 - internal semiconducting layer; 6 - conductor wire.

Expert System for Predictive Diagnosis (1) Principles and Knowledge Base

21

During operation, under the action of physical-chemical factors in the environment, in the cable occurs corrosion phenomena. These phenomena provoke degradations of the outer covering which further propagate to the active conductor. In time, the consequence of the corrosive action of environment is the emergence of possibilities of electrical undesirable interactions with the environment. These interactions, called currently faults, provoke the destructive effects of electricity, mainly, melting and evaporation of the materials involved. The faults occurrence means the interruption of power supply. The time duration of the interruption comprises the fault identification and location, the period of intervention preparation and the duration to eliminate the fault. In this duration, it is required, if possible, that the beneficiary, the consumer of electric power to be supplied from alternative sources, most often by switching other electrical functional subnets. It should be added that power networks are comprised of segments of cables that make up subnets between power substations. The discussion of this paper is limited to the power end consumers and suppliers groups, residential or industrial. It is to remember that maintenance of a network of power cables assumes high costs, both in terms of identification and location fault, and from the point of view of intervention. 2.3

Monitoring the Faults in Underground Power Cables

Hereinafter, the treatment of faulty underground power cable relates to the fault detection and location. The two actions are complementary and necessary. Fault detection requires knowing all its features. From this point of view is talking about short circuits and interruptions. In both cases the normal operation ceases but the consequences are regarding the rest of the network where the fault occurs. Locating the fault involves establishing the spatial coordinates of the place where it occurred. Since the underground power cables are not directly accessible, monitoring of their technical condition is quite always via indirect methods and techniques. Of practical interest particularly enjoys the predictive diagnosis of underground power cables. Predictive diagnostics [3,4] allows early fault detection and location. Such diagnosis is done by using some information on operation and maintenance, mainly resistances and voltages that can be measured within the network. The main parameters and indicators that enable predictive diagnostics are those given in Table 1. Analysis of this information will be carried out in relation with the destination, with the way to obtain, with interpretation and manipulation manner, with representation mode.

22

D. Isoc

Table 1. Knowledge associated to power cables and manner to treat the information inside the expert system.

(Continued)

Expert System for Predictive Diagnosis (1) Principles and Knowledge Base

23

Table 1. (Continued)

Note: In relationships in Table 1 was used the license that by [n] was symbolized the value of the nth row of the table.

3 3.1

Implementation Details Regarding the Expert System Describing the Knowledge

A first primary analysis of the available information indicates that this information: (a) is non-homogeneous; (b) it is diverse; (c) it comes from multiple sources and requires preliminary processing. Another characterization of available information is considering its purpose. The available information is used to ensure the prediction of life expectancy of underground power cables. Ensuring the life of the cable corresponds with an indirect anticipation of the moment of the next fault. 3.2

Principle of Using the Knowledge

The lack of homogeneity of the information underlying the decision-making process regarding the life of power cable requires the use of an expert system. Its main functions are: (i) Gathering primary information on a power cable according the template and value domains in Table 1 with examples of some situations as in Table 4. (ii) Saving of parameters of the cable introduced at operators option. (iii) Printing the information for a cable introduced into the system together with the conclusion on its state.

24

D. Isoc

The expert system involves a set of interactions that will be detailed in relation to the activities developed in cooperation with the human operator. Expected life of the cable is determined in relation to a set of values of predicted influence factors. As a rule, these influence factors parameters are obtained in relation to the nearest cable model found in the knowledge base and to the value of reference influence factor that the parameter has on the life of the cable: Real f actor =

Real parameter × Ref erence f actor Ref erence parameter

(1)

The expert system is based on the principles set out in [4] and [3] respectively. This requires the establishment of the implementation details to ensure operation within the expected limits. The expert is prepared to use the information in Table 1. This information corresponds to the used model. The same information can be provided by the user as a result of his relevant measurements. From Table 1 it follows how the designer provides connections between values and ways of interaction with the operator as in Table 2. In the information framework as in Table 1, there are a number of situations where the nature of the information should be specified because it must be validated during operation. A designation of some of these situations is given in Table 3. The established relations are those that customize the processing manner with the operation and ensure the specificity of the interaction way of implementing of expert system. What is specific in this step refers to the fact that: Table 2. Notations of the ways to process the information. Notation Explanations regarding the associated processing way A

Default value is zero

B

One inserts by the operator as numerical value and one validates according to accepted value field. One inserts both when the operator opts for introducing a reference cable and when the operator wants to introduce a cable to be analyzed

C

One inserts by the operator as numerical value and one validates according to accepted value field. One inserts only when the operator opts for introducing a reference cable

D

One inserts by the operator by selection

E

Automatically selected by the system according the value selected by the operator

F

One finds automatically using the associated and explained relations in Table 1

Expert System for Predictive Diagnosis (1) Principles and Knowledge Base

25

Table 3. Symbols used to codify the data types. Symbol Data type DA

Alpha-numerical string

DB

Real, positive value

DC

Real value

DD

Binary, logical value

(i) That the information on underground cables is comprised of two distinct sets: (a) a set of information obtained directly by measurements or factual findings; (b) a set of information relating to the influence factor associated to measured parameter during the normal operation of underground cable. (ii) That it is established how to get information on the analyzed cable, i.e. from human operator by direct input and validation for numerical information or from human operator by selecting, for other kind of information. In special situations, there is a possibility of open format, especially for information that is not relevant to processing but only for identifying the work cases. (iii) That the information entered by the operator is operated by intermediate information associated to influence factors on the normal life of the cables. In fact, this information is the expert system work manner and assumes the included expertise that can not be described directly in the working mechanism of expert system. In Table 1 there is a large amount of information to be processed. Much of this information is entered by a human operator in two different situations. The first situation is where the human operator wants to introduce in the knowledge base of expert system a new piece of knowledge. The knowledge identifies to a new cable. The second situation is where the human operator introduces the situation of a cable to be analyzed and treated by means of information known by the expert system. In all cases, decisive is the way to build the cable model that is given by means of the set of influence factors acting together with the characteristic variables, known as attributes, on the value of the life of the cable. In analyzing the data in Table 1 it is necessary to point out that here, in the white fields, is the measurable information or information that can be collected and each is provided with a weighting factor. Weighting factors present in the gray cells are defaults as value domain and they emphasize the importance that the informed user, the human expert, gives to the variables associated to the attributes of knowledge face to the possible complex relation to get the life of the cable. In preparing the connection between human operator and expert system important is to point the manner of treatment of information. This responsibility

26

D. Isoc Table 4. Particular value of parameters. Symbol Particular value VA

Sample 1 PT Iugoslaviei/ ST South fS

VB

{A2Y SY, A2xS2Y, A2XSR2Y, A2XSRY − B}

VC

(a) buried, directly in soil; (b) buried in protective coating

VD

For: (a) buried, directly in soil the parameter has the real value 50, and for (b) buried in protective coating the associated value is 1

is determined to implement or allocate a specific value. This is specific to expert system implemented in connection with the information in Table 1 and detailed further in Table 2. In the information that appears in Table 1 exists a number of situations where the nature of information should be specified because it will be validated during operation. A designation of some of representative cases is given in Table 3. In certain situations in Table 1 shall apply some notations to be used when the parameter are difficult to introduce into a simplified table, in which case are allowed notations as in Table 4. 3.3

Implementing the Working Algorithm

For best results in predicting the cable life expectancy, each cable is described by a table which has elements associated with the relevant information for describing the cable, but also for its operation. It is assumed that one works with a reference cables Cai that have the parameters {pai1 , pai2 , pai3 , pai 4} and the value of the cable life expectation is pai0 . In relation to this reference cable is considered another cables for which were achieved measurements where from the parameter set is {pax1 , pax2 , pax3 , pax4 }. It is seeking to determine the predicted lifetime px0 of the given cable. The assumption implied by the prediction method is that each parameter intervenes on the cable life expectation by an influence factor {wi1 , wi2 , wi3 , wi4 }, value resulting from appreciation heuristic. This is the reason why all the information on a reference cable Cai , i.e. its parameters, paij , j = 1 · · · m, is entered into the reference data base as in Table 1. Be a case where it is a Cai reference cable with a number of m = 4 parameters. In this case the associated parameter structure is formed with the values: Cai (1, 1) = pi1 ; Cai (1, 2) = wi1 ; Cai (2, 1) = pi2 ; Cai (2, 2) = wi2 ; Cai (3, 1) = pi3 ; Cai (3, 2) = wi3 ;

Expert System for Predictive Diagnosis (1) Principles and Knowledge Base

27

Cai (4, 1) = pi4 ; Cai (4, 2) = wi4 ; Cai (5, 1) = pi0 ; Cai (5, 2) = 1; In order to achieve a predictive diagnosis, the cable Cax is considered. For this cable, the life expectation is seeking and that one builds its own array of values: Cax (1, 1) = pax1 ; Caxi (1, 2) = wx1 ; Cax (2, 1) = pax2 ; Caxi (2, 2) = wx2 ; Cax (3, 1) = pax3 ; Caxi (3, 2) = wx3 ; Cax (4, 1) = pax4 ; Caxi (4, 2) = wx4 ; where the cable life and operating time is kept in the table raw Cax (5, 1) = pax0 which corresponds to Cax (5, 2) = wx and where wx =

wx1 + wx2 + wx3 + wx4 wi1 + wi2 + wi3 + wi4

(2)

pax0 = pai0 × wx

(3)

and The weights {wx1 , wx2 , wx3 , wx4 } can be determined in two ways: – by reporting the parameters of cable to the parameters of reference cable as: wx1 =

pax1 pai1

pax2 pai2 pax3 = pai3 pax4 = pai4

(4)

wx2 =

(5)

wx3

(6)

wx4

(7)

– by other computing relationships into which, for example, one finds the relationships between measured parameters as in the manner: wx1 = f1 (pax1 , pax2 , pax3 , pax4 )

(8)

wx2 = f2 (pax1 , pax2 , pax3 , pax4 )

(9)

wx3 = f3 (pax1 , pax2 , pax3 , pax4 )

(10)

wx4 = f4 (pax1 , pax2 , pax3 , pax4 )

(11)

28

D. Isoc

Table 5. The arrays, side by side, of the parameter of reference cables and of analyzed cable. ca1

ca2

ca3

ca4

p11 w11 p21 w21 p31 w31 px1 wx1 p12 w12 p22 w22 p32 w32 px2 wx2 p13 w13 p23 w23 p33 w33 px3 wx3 p14 w14 p24 w24 p34 w34 px4 wx4 p15 w10 p25 w20 p35 w30 px5 wx0

The above situation corresponds to the case where the reference base has a single cable or a search has been found that the closest to the examined cable of Cax is the i - th cable, i.e. Cai . Now it considers that the reference cable base, with the same number of parameters, comprises the cables Ca1 , Ca2 , Ca3 to which the analyzed cable Cax is reported in order to predict its life expectation. The parameter array is being built in parallel as shown in Table 5. According to the principle of prediction, it must be found the Ci cable that is most similar to the analyzed cable, Cx in terms of all description parameters. Once the parameters of analyzed cable i.e. pax1 , pax2 , pax3 , pax4 were determined by measurement or by other way, it follows the stage of selecting the most similar cable. The choice is made by a synthetic indicator resulted for each comparison basis. For reasons of validity, one uses the mean square error and then it follows: – for the Cax and Ca1 cables:  E1 = 0.25 × [(p11 − px1 )2 + (p12 − px2 )2 + (p13 − px3 )2 + (p14 − px4 )2 ] (12) – for the Cax and Ca2 cables:  E2 = 0.25 × [(p21 − px1 )2 + (p22 − px2 )2 + (p23 − px3 )2 + (p24 − px4 )2 ] (13) – for the Cax and Ca3 cables:  E3 = 0.25 × [(p31 − px1 )2 + (p32 − px2 )2 + (p33 − px3 )2 + (p34 − px4 )2 ] (14) With the degree of similarity between the analyzed cable and each reference known, the selection process is capable of determining the most similar cable, that is of the cable with the i-index and so that the value of synthetic index of similarity is maximized: i = arg mini (E1 , E2 , E3 )

(15)

Expert System for Predictive Diagnosis (1) Principles and Knowledge Base

29

We can identify two situations: – If there is only one cable to the one observed, be it Ca2 . In this situation, the predicted life expectation of analyzed cable is given in terms of the influence factors of the parameters on its and the global influence factor of the analyzed cable wx : wx1 + wx2 + wx3 + wx4 wx = (16) w21 + w22 + w23 + w24 concretely pax0 = pa20 × wx (17) – If there are more similar cables with the analyzed cable, be them Ca1 and Ca3 . In this situation, cable life expectation will result as a weighted average pax0 = pa10

wx1 + wx2 + wx3 + wx4 wx1 + wx2 + wx3 + wx4 + pa30 w11 + w12 + w13 + w14 w31 + w32 + w33 + w34

(18)

A graphical representation of parts of the knowledge associated with the information characterizing a cable is given in Fig. 2. In Fig. 2, the information on the monitored cable is ordered in a part called Cxi context of knowledge Pi , and the part called the actual knowledge, Ki .

User

Pi1=[1]

Pi2=[3]

Cxi

f1

1

f2

Cxi

2

↔ Pi17=[33]

f17 17

Ki

Pi

Fig. 2. Information about the cable in the structure of a knowledge piece.

The components of knowledge pieces are related to the information in Table 1. Through the context of a piece of knowledge is meant all information substantiating the knowledge. In another form, they are also called premises. To a set of premises one associates the knowledge itself.

30

D. Isoc

Table 6. The algorithm for determining the predicted lifetime of the analyzed cable with respect to a base of reference similar cables.

Expert System for Predictive Diagnosis (1) Principles and Knowledge Base

31

Context and knowledge are the effect of a life experience or an experience verified by an expert or, generally, of a connoisseur. Operation of an expert system dedicated to prediction allows that for a new context, to anticipate the conclusion or the knowledge that can be reached. The associated algorithm is given in Table 6. From the algorithm shown in Table 6 it can be ascertained that: (i) each known cable is a piece of knowledge; (ii) the pieces of knowledge have many component types; (iii) the component parts of the piece of knowledge are of many types: nonnumerical data, experimental determined numerical data, numerical data determined by calculating using indirect ways; (iv) each component part of each piece of knowledge intervenes in reality describing by means of a weighting or influence factor. The weighting factors are unique for a given knowledge base and they are depending of the accumulated expertise of its users. (v) each piece of knowledge enters in the inference result in term with the similarity degree of analyzed cable with the given piece of knowledge. A scheme for an entirely functional expert system dedicated to predictive diagnosis in given in Fig. 3. SCU

S1 S2

SU

AU

Cx1

Cx2

K1 P1 KB

K2

Sn

Cxn

P2

Kn

Pn

P0

P0 IE Cx0

Cx0

?

K0

Fig. 3. The block scheme of an expert system for predictive diagnosis.

32

D. Isoc

Here are recognizable the modules and functional connections of the expert system. The knowledge base as KB is one that comprises an expertise organized as a set of n confirmed experiments described by sheets constructed as in Table 1. This expertise matches a set of cables for which is known the characterization information and the life of operation. Each piece of knowledge P 1, P 2, · · · consists of a context, Cx1, Cx2, · · · and an associated knowledge K1, K2, · · · . The purpose of the expert system is to determine the K0 knowledge that could correspond to known context, Cx0 . In order to do that, one seeks to select the situations in knowledge base KB that are most close with the given problem and so, filling the piece of knowledge, P0 . The process of determining follows the course of the selection process which is done in a selection unit, SU under the control of a selection control unit SCU . Through selectors S1, S2, · · · , Sn the context of knowledge pieces is compared with the context of the problem knowledge piece, P0 . At the end of selection, an enabling unit, AU allows to these s leq n knowledge pieces to access an inference machine, IE. Through inference algorithm described in Table 6 the inference engine predicts the expectation life the presented problem presented and complements the knowledge piece P0 with the knowledge K0.

4

Concluding Remarks

In this paper one argues the achievement principles of a knowledge base of an expert system dedicated to predictive diagnosis, that is the evaluation of the life expectation of a power cable working underground an is subjected to corrosion. It is ascertained that modeling of underground power cable requires a significant amount of non-homogeneous information of individual natures. The knowledge base is built of cables about one knows how they have evolved over time and operate or the accidents that have affected. It is anticipated that the manner of using of knowledge will be by means of an expert system that works by making a weighted average of numerical information representing the available knowledge. The final value of the predicted cable life expectation requires knowledge as numerically weighted on values with associated significance following procedures or experiences. Each cable known formerly the prediction is one knowledge piece and contributes to determining the value of life expectation. The structure and operating manner of designed expert system are as the standard, but the inference algorithm is a specific one and works with specific knowledge pieces.

Expert System for Predictive Diagnosis (1) Principles and Knowledge Base

33

References 1. Gazdzinski, R.: Cable and method of monitoring cable aging. US Patent 5902962 (1997) 2. Goodman, B., Guthrie, G., Starke, W., Stuecheli, J., Williams, D.: Data processing system and method for predictively selecting a scope of broadcast of an operation utilizing a history-based prediction. US Patent 7444494 (2005) 3. Isoc, D.: Diagnosis, and fault detecting structures in complex systems. In: Study and Control of Corrosion in the Perspective of Sustainable Development of Urban Distribution Grids - The 2nd International Conference, Miercurea Ciuc, Romania (2003) 4. Isoc, D., Ignat-Coman, A., Joldi¸s, A.: Intelligent diagnosis of degradation state under corrosion. In: Arioui, H., Merzouki, R., Abbassi, H.A. (eds.) Intelligent Systems and Automation, pp. 383–391. Melville, New York (2008) 5. Le Gressus, C., Faure, C., Bach, P., Blaise, G.: Process for the reduction of breakdown risks of the insulant of high voltage cable and lines during their agging. US Patent 5660878 (1995) 6. Navratil, R.: Monitoring and fault detection in dynamic systems. US Patent 7421351 (2006) 7. Schweitzer Jr., E.: Housing including biasing springs extending between clamp arms for cable mounted power line monitoring device. US Patent 5180972 (1992) 8. Soma, K., Kotani, K., Takaoka, N., Ikeda, C., Marumo, M.: Method for diagnosing an insulation deterioration of a power cable. US Patent 4980645 (1990) 9. Szatmari, I., Lingvay, M., Tudosie, L., Cojocaru, A., Lingvay, I.: Monitoring results of polyethylene insulation degradability from soil buried power cables. Rev. Chim. 66(3), 304–311 (2015) 10. Thor, T., Pellerito, B.: Clutch fault detection. US Patent 7421326 (2004) 11. Yagi, Y., Tanaka, H.: Method of diagnosing deterioration of the insulation of an electric power cable. US Patent 6340891 (1999)

Expert System for Predictive Diagnosis (2) Implementing, Testing, Using Dorin Isoc(B) Technical University of Cluj-Napoca, Cluj-Napoca, Romania [email protected]

Abstract. Once the conceptual part defined and established, developing an expert system assumes steps of engineering in the context of practices and procedures. Obviously, the already achieved expert system requires adjustments and refinements involving a significant effort. There are given principles to design the man-machine interface and the operating algorithm. Follows specifying the principles of testing technologies and then build it, together with the test data. The work exemplifies two of verification tests that prove also how to use the expert system. Keywords: Expert system · Underground power cable · Knowledge base · Knowledge piece · Approximate reasoning · Man-machine interface · Test technology · Test data

1

Introduction

In general, the development of an expert system is treated mostly by mathematical grounds and of treating of knowledge. Rarely, details as man-machine interface and system testing are considered as objects worthy of attention for scientific research. We take advantage of the fact that expert system for diagnosis of electrical underground power cables is an application sufficiently specialized to be described fully, including the ways of construction of the solution its manmachine interface and how it was built testing technology together with the test data set. Among the applications that were the basis of documenting, there is an adaptive application [6] in the sense that the interface is built on the latest operated applications. The diagnostic expert system is characterized by the need to complete the entire set of information and then, by activating a simple inference engine. In this way such a solution is not justified. Comparative analysis of [9] shows a big difference between interfaces made in universities and the commercial products. It is noted especially the complexity of products made in universities. It also noted the most abundant use of techniques and details in interfaces achieved in universities, such as graphic objects, mouse, c Springer International Publishing AG 2018  V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8 4

Expert System for Predictive Diagnosis (2) Implementing, Testing, Using

35

screen, text editing, use of hypertext and gesture recognition. One explanation might be the lack of a pertinent, realistic vision on the effective use of a manmachine interface. Is also noteworthy that, despite the growing number of expert systems in recent years, research reported in graphical user interfaces exist, especially in 80–90 years [5,7,10]. By intent, it is distinguished paper [8] where it one put a synthesis of conceptual issues that could form the basis for human-machine interfaces. However, the main conclusion is that which would be required depends on the design methodology which does not really exist. In the absence of methodologies to build man-machine interface for application to develop expert systems, the author believes that it is useful to approach the issue and solve it the most explicitly. In this context, the first conclusion of the author is that for a rational use and within the limits of its intended destination as diagnostic expert systems, the man-machine interface must be designed by one who knows both the knowledge base and the way of thinking of the user. It also retains the assembly ergonomicseffective operation need-algorithmic description. Together with the man-machine interface it is approached the developing of testing technology and the developing of test data set. In the first part of the paper are given details of man-machine interface, including its specification. In a second part are given the test principles of expert system and further are introduced two relevant case studies for testing technology and test data set choice. Finally, one concentrates the conclusions of this approach to build the expert system for predictive diagnostics.

2

Details on Specification and Building a Man-Machine Interface for an Expert System

In general, between the expert system and the human operator there is a specific action. The interaction is guided by the man-machine interface. It is profoundly wrong to consider that an effective interface can be derived from a general purpose interface or the human-machine interface system should enjoy all the amenities of a general purpose product-program. 2.1

General Functions

The interface is a constructive detail of the expert system that gives it accessibility and, finally, the acceptability of the beneficiaries. In this situation, the interface is built having in mind hypothesis of industrial implementation, but the product is made in minimal technological conditions to ensure the operation according to the stated principles. Functions which the interface has to provide are: i. Reception and validation of information that characterizes the cable to be analyzed. In order to receive the requested knowledge is defined the information to be introduced by the human operator as in Table 1.

36

D. Isoc

ii. The possibility that the operator can simply act to attain the expert system goal without having any special training. iii. The possibility of software to be easily tested, especially since it is an expert system operating with complex knowledge. iv. The possibility that the solution offered by the expert system to be justified in the sense indicated by the element of knowledge base that was considered like closest to the present analyzed case. The expert system interacts with the operator by means of several actions: i. entering data for the purposes of characterizing analyzed cable by measurable parameters; ii. operative activation of commands to achieve desired actions; iii. input date validation; iv. reporting the parameters of analyzed cable and of resulted solution, respectively of predicted lifetime. In Table 2 are defined selectors that should be provided for the expert system dedicated to analysis and life prediction of underground power cables. In Table 3 are described the effects of commands and how they interact with other possible actions. 2.2

Building the Interface of Expert System

Expert system interface is designed to implement defined commands. The principle that it respects the interface is that its users have enough freedom of action in a defined framework of limitations and safeguards. In Fig. 1 is given the expert system’s main board. It is ascertained that there is a part for working parameters. It was chosen that this sub-screen have dual utility. It is to remember that maintenance of a network of power cables assumes high costs, both in terms of identification and location fault, and from the point of view of intervention.

3

Testing and Using the Expert System

Testing an expert system is a necessity before it is placed in service. There is no single test technology. In the literature, which is recommended, in different forms [1–4], is going through all routes leading from premises to conclusions. Such exhaustive testing is suitable for expert systems that work on a limited and predefined knowledge base. In this case it is a different kind of expert system. First, the knowledge base is not limited but it enriches by means of accumulated experience. Accumulation of knowledge can be done from the outside from accumulated experiences or by including situations already considered and solved.

Expert System for Predictive Diagnosis (2) Implementing, Testing, Using

New job

Operating

A N A L Y S I S

Validate

Activate prediction

Abandon

37

U P D A T I N G

Validate Catalogue inspecting Cable saving Cable removing

Exit

Fig. 1. Placement of commands on the operating sub-panel.

When tests an expert system it should keep in mind that behind the results are more knowledge together with the rules of use. Components of knowledge pieces are at least of a three categories: (a) numerical with direct assignment, (b) calculated based on predefined account relationships and (c) other, basically logical. It is noteworthy that the component parts of knowledge pieces have not identical relations with the lifetime of the analyzed cable. This makes the testing technology still more complicated. It is noted that verification test should be performed with a large number of cases, that reference cables and checking the components parts of knowledge to be ‘practiced’, that is to have many values. 3.1

Organizing of Testing

In these circumstances, the test is regarded by examples. In a superficial manner, testing is exemplified by construction the tests as below: Test 1. The cable has identical parameters with the parameters of first of the analyzed reference cables of the knowledge base. It may be noted that there are three tables that relate to cable analyzed (Table 1) which includes the solution of expert system, the knowledge base describing the set of reference cables (Table 2) and a table that allows the evaluation of details that refers to the joining of analyzed cable and the reference cable.

38

D. Isoc Table 1. Test 1 – reported information of analyzed cable.

No

Parameter

Lower Sample

[1]

Cable type

A2YSY

Upper

[2]

IF of the cable type on the CLE

0

[3]

Laying technology

A

[4]

IF of laying technology on the CLE

1.00

50.00

50.00

[5]

Cable section [mm2 ]

10.00

240.00

500.00

[6]

IF of the cable section on the CLE

[7]

Length of cable segment [m]

[8]

IF of the cable length on the CLE

[9]

Insulation resistance, [MΩ] − &5 kV

0.01

0.01

1,000,000

[10] IF of the insulation resistance on the CLE

1.00

86.95

100.00

[11] Shield/soil insulation resistance, [MΩ] − &5 kV

0.00

0.10

300.00

[12] Shield/soil insulation resistance IF on CLE

1.00

1.00

100.00

0.00 20.00

1,880.00 10,000 0.00

[13] Shield electric continuity

Yes

[14] IF of shield electric continuity on CLE

1.00

[15] Dielectric loss tangent (tgδ)

0.0005 1.00

100.00

1.00

100.00

[16] IF of dielectric loss tangent on CLE

1.00

100.00

100.00

[17] Soil resistivity [Ω m]

1.00

4.50

100.00

[18] IF of soil resistivity on CLE

1.00

1.00

100.00

[19] Soil acidity [pH]

2.00

5.50

12.00

[20] IF of soil acidity on the CLE

1.00

31.00

100.00

[21] Stray current density [A/m2 ]

0.00

0.03

1.00

[22] IF of stray currents density on the CLE

1.00

3.67

100.00

[23] Applied sleeves number

0

24

30

[24] IF of applied sleeves number on CLE

1.00

1.00

100.00

[25] Mean cable load [A]

5.00

1,200.00 3,000.00

[26] IF of mean cable load on CLE

10.00

37.86

100.00

[27] End 1 shield/soil voltage [VCu /4CuSO ]

−1.00

0.15

0.50

[28] IF of end 1 shield/soil voltage on CLE

1.00

100.00

100.00

[29] End 2 shield/soil voltage [VCu /4CuSO ]

−1.00

0.02

0.50

[30] IF of end 2 shield/soil voltage on CLE

1.00

52.69

100.00

[31] Year of cable’s laying

1,930

1,975

cy

23.00

50

[32] Estimated life expectation for the known cable, in years 0.00

Test 2. This test is designed to test the functioning of the expert system using external, reliable information. This information, resulting from expert judgment says that an essential factor of corrosion, in identical conditions, plays the mean cable load. 3.2

Test Results

The complexity of dependencies that underlie of built expert system raises serious test questions.

Expert System for Predictive Diagnosis (2) Implementing, Testing, Using

39

Table 2. Test 1 – database contents to be used. No

Parameter

Ex.1

Ex.2

Ex.3

Ex.4

[1]

Cable type

A2YSY A2YSY

[2]

IF of the cable type on the CLE

0.00

0.00

0.00

0.00

[3]

Laying technology

A

A

A

A

[4]

IF of laying technology on the CLE

50.00

50.00

50.00

50.00

[5]

Cable section [mm2 ]

240.00

240.00

300.00

300.00

[6]

IF of the cable section on the CLE

0.00

0.00

0.00

0.00

[7]

Length of cable segment [m]

1,880

1,880.00 400.00

400.00

[8]

IF of the cable length on the CLE

0.00

0.00

0.00

0.00

[9]

300.00

A2YSY A2YSY

Insulation resistance, [MΩ] − &5 kV

10.00

10.00

300.00

[10] IF of the insulation resistance on the CLE

87.00

87.00

86.00

87.00

[11] Shield/soil insulation resistance, [MΩ] − &5 kV

0.01

0.01

3.15

0.21

[12] Shield/soil insulation resistance IF on CLE 100.00

100.00

16.00

38.00

[13] Shield electric continuity

Yes

Yes

Yes

Yes

[14] IF of shield electric continuity on CLE

100.00

100.00

100.00

100.00

[15] Dielectric loss tangent (tgδ)

1.00

0.99

0.08

0.07

[16] IF of dielectric loss tangent on CLE

100.00

99.46

8.91

7.42

[17] Soil resistivity [Ω m]

4.50

4.50

11.10

11.10

[18] IF of soil resistivity on CLE

1.00

1.00

1.00

1.00

[19] Soil acidity [pH]

5.50

5.50

6.50

6.50

[20] IF of soil acidity on the CLE

31.00

31.00

11.00

11.00

[21] Stray current density [A/m2 ]

0.03

0.03

0.06

0.06

[22] IF of stray currents density on the CLE

3.67

3.67

6.64

6.64

[23] Applied sleeves number

24

24

3

7

[24] IF of applied sleeves number on CLE

1.00

1.00

1.00

1.00

[25] Mean cable load [A]

1,200

1,200

1,300

1,300

[26] IF of mean cable load on CLE

37.86

67.24

41.53

41.53

[27] End 1 shield/soil voltage [VCu /4CuSO ]

0.15

0.15

−0.20

0.01

[28] IF of end 1 shield/soil voltage on CLE

100.00

100.00

37.59

48.71

[29] End 2 shield/soil voltage [VCu /4CuSO ]

0.02

0.01

−0.27

−0.12

[30] IF of end 2 shield/soil voltage on CLE

52.69

47.20

5.37

17.97

[31] Year of cable’s laying

1975

1975

2005

2005

[32] Estimated life expectation for the known cable, in years

23.00

12.00

45

26

In Test 1 is checking the reproduction of a known solution. It is ascertained that when the parameters of analyzed cable are identical to those of a cable of the set of reference cables, the lifetime of analyzed cable is the same of the reference situation of Table 3. In the case of Test 2 one checks a particular solutions, quasi-known.

40

D. Isoc Table 3. Test 1 – details of monitoring the right operating of the expert system.

No

Parameter

Sample Reference

[1]

Cable type

A2YSY A2YSY

[2]

IF of the cable type on the CLE

0.00

[3]

Laying technology

A

A

[4]

IF of laying technology on the CLE

50.00

50.00

[5]

Cable section [mm2 ]

240.00

240.00

[6]

IF of the cable section on the CLE

0.00

0.00

[7]

Length of cable segment [m]

1,880

1,880.00

[8]

IF of the cable length on the CLE

0.00

0.00

[9]

Insulation resistance, [MΩ] − &5 kV

10.00

10.00

0.00

[10] IF of the insulation resistance on the CLE

87.00

87.00

[11] Shield/soil insulation resistance, [MΩ] − &5 kV

0.01

0.01

[12] Shield/soil insulation resistance IF on CLE

100.00

100.00

[13] Shield electric continuity

Yes

Yes

[14] IF of shield electric continuity on CLE

100.00

100.00

[15] Dielectric loss tangent (tgδ)

1.00

1.00

[16] IF of dielectric loss tangent on CLE

100.00

100.00

[17] Soil resistivity [Ω m]

4.50

4.50

[18] IF of soil resistivity on CLE

1.00

1.00

[19] Soil acidity [pH]

5.50

5.50

[20] IF of soil acidity on the CLE

31.00

31.00

[21] Stray current density [A/m2 ]

0.03

0.03

[22] IF of stray current density on the CLE

3.67

3.67

[23] Applied sleeves number

24

24

[24] IF of applied sleeves number on CLE

1.00

1.00

[25] Mean cable load [A]

1,200

1,200

[26] IF of mean cable load on CLE

37.86

37.86

[27] End 1 shield/soil voltage [VCu /4CuSO ]

0.15

0.15

[28] IF of end 1 shield/soil voltage on CLE

100.00

100.00

[29] End 2 shield/soil voltage [VCu /4CuSO ]

0.02

0.02

[30] IF of end 2 shield/soil voltage on CLE

52.69

52.69

[31] Year of cable’s laying

1975

1975

[32] Estimated life expectation for the known cable, in years 23.00

23.00

By the construction of this situation, for the same knowledge base the lifetime value is significantly influenced as in Table 4. Here only the significant rows are presented.

Expert System for Predictive Diagnosis (2) Implementing, Testing, Using

41

Table 4. Test 2 – details of monitoring the right operating of the expert system. No

Parameter

···

···

Sample Reference ···

···

[25] Mean cable load [A]

2,500

1,200

···

···

···

[32] Estimated life expectation for the known cable, in years 12.69

··· 23.00

The above two situations are part of the tests battery that was built to validate the achieved expert system. From this point of view, it is considered that the assessment succeeded. Regarding the information in Table 1 is to remember some details: i. The introduction of information on the analyzed cable involves both objective information, and entering the values of influence factors. ii. Information on all parameters of cable, inclusively of influence factors, may be partially in the sense that the expert system assigns defaults. In this way the system can cope with imprecise information. The existence of influence factors covers that knowledge of the expert that is doubtful through ignorance regarding the involved physic-chemical phenomena. Different values of influence factors in Table 1 indicate that referred specimens of cables are coming from different sources. There, the specialists who assessed them found that they could be the values that they are able to support them. The size of reference cable set is a guarantee of prediction quality. It is obvious that the predicted value is among the available knowledge pieces. Clearly the highlight of the weight to affect each knowledge piece in the final outcome is difficult. By the nature of the tasks undertaken testing, the research that might be necessary will be avoided. Test 2 performed mainly based on information from Table 4, makes a check that covers both the theoretical used basis and how to implement a system expert. Construction of the test is based on the assumption that doubling the cable mean load should lead to a half of lifetime of the cable if other parameters are not changed. Tables 3 and 4 have the role of working tools. Together with Table 1 they form the testing technology of the expert system. The task of designing these documents devolve upon the project group. Despite many aspects of fundamental research, these documents are part of professional product documentation and once built the product, it must be sent to the end user. Testing will not be completed with technical reception of expert system. The testing must be repeated after a period of time agreed with beneficiary and must be extended to newly entered knowledge pieces whether they are analyzed cables whether they are the results of laboratory research or experimental results.

42

D. Isoc

It requires periodic statistical checking regarding the predictions and results confirming the favorable results. Testing can be completed with achievement of graphics with intermediate possible values (Table 5). Table 5. Knowledge associated to power cables and manner to treat the information inside the expert system. No

Explanation about the information Name heading comments

Lower

Field to take the value Upper

[1]

Cable location

[2]

Cable type

Selector S1

[3]

Laying technology

Selector S2

[4]

Cable section [mm2 ]

10

500.00

[5]

Length of cable segment [m]

20

10, 000

[6]

Insulation resistance, [MΩ] − &5 kV

10

10

[7]

Shield/soil insulation resistance, [MΩ] − &5 kV

0.01

30

[8]

Shield electric continuity

[9]

Dielectric loss tangent (tgδ)

Selector S3 0.00005 Selector S3

1.00

[10] Soil resistivity [Ω m]

1.00

100.00

[11] Soil acidity (pH)

2.00

12.00

[12] Stray currents density [A/m2 ]

0.00

1.00

[13] Applied sleeves number

0

[14] Mean cable load [A]

5

Selector S4

30 3, 000

[15] End 1 shield/soil voltage [VCu/CuSO4 ] −1.00

0.50

[16] End 2 shield/soil voltage [VCu/CuSO4 ] −1.00

0.50

[17] Year of cable’s laying

1930

Selector S5

cy

[18] Estimated life expectation for the known cable, in years

0

Selector S6

50

On the one hand sub-screens of work parameters is used for inspection, maintenance and updating of the knowledge base. On the other hand, during the analysis, the same sub-screens will provide the final result on the position of the predicted lifetime (Table 6). In Fig. 2 one gives the operating sub-screen. This is divided in three key areas. The first area is the introduction of a new reference cable or advance under the existing cable base during maintenance. The other two areas correspond to the analysis commands for updating and maintaining the database of reference cables. The third area is for mandatory commands i.e. abandon of last activated command and exit command (Table 7).

Expert System for Predictive Diagnosis (2) Implementing, Testing, Using

43

Table 6. Defining selectors of man-machine interface of the system expert. Description of interface selectors S1 - cable type selector Values to be selected: A2YSY, A2xS2Y, A2XSR2Y, A2XSRY-B Default state:

A2YSY

Notes:

Does’nt exist

S2 - laying technology selector Values to be selected: {Buried directly in soil, buried in protective coating} Default state:

Buried directly in soil

Notes:

Does’nt exist.

S3 - electric continuity of cable shield Values to be selected: {Yes, No} Default state:

Yes

Notes:

Does’nt exist.

S4 - selector of sleeves number Values to be selected: [0, 30] Default state:

0

Notes:

Does’nt exist.

S5 - cables laying year selector Values to be selected: 1930, 1931, 1932, · · · cy Default state:

(cy 5)

Notes:

With cy is denoted the current year.

S6 - selector of known operating time Values to be selected: {0, 1, 2, · · · 50} Default state:

5

Notes:

Selector is active only for the parameters updating stage

ElCabExpert

Expert system for power cables life prediction

Work parameter screen

Operating screen

Fig. 2. Organizing the board of expert system.

It is to emphasize that one takes into account that the two work commands regarding analyze and updating are not commands on the same hierarchical rank.

44

D. Isoc Table 7. Defining man-machine interface commands of the system expert.

Command/description New job - Default command, active in initial state - Allows the acting of commands to set the work mode, that is Analysis for the mode of Power cable analysis, and respectively Updating for the work mode Cable basis updating and also of the general commands Abandon and Exit - Do not allow parameter fields filling with values until the activation - Allows repeated activation of command - Bring the system in the specific state of operating mode Cable analysis - Do not allow activating of specific commands of Activate prediction, and/or Cable saving and Catalog inspection Analysis - Default command, active in initial state - Defines the work mode Power cable analysis - Allows the activating of specific commands Activate prediction and of general commands Abandon and Exit - Do not allow the activating of specific commands of Cable base updating work mode namely Cable saving and Catalog inspection - Allows parameters entering for the analysis cable but do not allow the fulfill of the field of known lifetime of the cable operating Estimated lifetime Updating - Default command, passive in initial state - Defines the work mode of Cable base updating - Allows the activating of specific commands Cable saving and Catalog inspection and of general commands Abandon and Exit - Do not allow the activating of specific command of Power cable analysis work mode namely Activate prediction - Allows parameters entering for the analysis cable, inclusively the fulfilling of the field of known lifetime of the cable operating estimated lifetime Power cable analysis/Activate prediction - Default command, active in initial state - Can be activated only successfully activated of Power cable analysis/Validate command - Replaced the inference engine and displays the estimate lifetime in its field Power cable analysis/Entering the parameter values /Abandon - Always asks for the acknowledgement to execution Do you want to abandon? (Yes/No) - Once confirmed, the system passes to the state activated by Analysis Cable updating/Entering the parameter values/Abandon - Always asks for the acknowledgement to execution Do you want to abandon? (Yes/No) - Once confirmed, the system passes to the state activated by Updating Exit - Always asks for the acknowledgement to execution Do you want to exit? (Yes/No) - Once confirmed the operator quit the expert system without any saving

Expert System for Predictive Diagnosis (2) Implementing, Testing, Using

45

This means that these commands can not be accessed by the other person than having administrator authority. This specificity is resolved by the existence of an access password. By recognizing the password, the operator can access distinctly analysis command or analysis and the updating commands.

4

Testing and Using the Expert System

Testing an expert system is a necessity before it is placed in service. There is no single test technology. In the literature, which is recommended, in different forms [1,3], is going through all routes leading from premises to conclusions. Such exhaustive testing is suitable for expert systems that work on a limited and predefined knowledge base. In this case it is a different kind of expert system. First, the knowledge base is not limited but it enriches by means of accumulated experience. Accumulation of knowledge can be done from the outside from accumulated experiences or by including situations already considered and solved. When tests an expert system it should keep in mind that behind the results are more knowledge together with the rules of use. Components of knowledge pieces are at least of a three categories: (a) numerical with direct assignment, (b) calculated based on predefined account relationships and (c) other, basically logical. It is noteworthy that the component parts of knowledge pieces have not identical relations with the lifetime of the analyzed cable. This makes the testing technology still more complicated. It is noted that verification test should be performed with a large number of cases, that reference cables and checking the components parts of knowledge to be ‘practiced’, that is to have many values.

5

Concluding Remarks

Once conceptually done, an expert system should be achieved as inference engine but also as human-machine interface. This interface comes to provide to users access to obtain right answers to questions. Achieved interface put together necessary functions and the facility of receiving information on the knowledge pieces that are then exploited. Since the users have different authority one creates the user with access to all the information, i.e. including the rights to update the knowledge base, and an ordinary user who has access only to solve key issues that prediction lifetime of underground power cables. Such information is used in development projects of the power supply network. It is grounded and exemplified the testing technology of achieved expert system.

46

D. Isoc

From the tests set that have been designed and implemented, in work are given two tests. The first test is the reproduction of the identical case and the second is to obtain predictable solution based on the information that comes from the expert in problems of corrosion.

References 1. Balci, O., Smith, E., O’Keefe, R.: Validation of expert system performance. Technical report TR-86-37, Department of Computer Science, Virginia Polytechnic Institute and State University (1986). eprints.cs.vt.edu 2. Bobrow, D., Mittal, S., Stefik, M.: Expert systems: perils and promise. Commun. ACM 29(9), 880–894 (1986) 3. Chang, C., Stachowitz, R.: Testing expert systems. Technical report (1988). ntrs.nasa.gov 4. Heckerman, D., Horvitz, E., Nathwani, B.: Toward normative expert systems: the Pathfinder project. Methods Inf. Med. 31, 90–105 (1991) 5. Hoc, J.: From human-machine interaction to human-machine cooperation. Ergonomics 43(7), 833–843 (2000) 6. Hoffberg, S., Hoffberg-Borghesani, L.: Ergonomic man-machine interface incorporating adaptive pattern recongnition based control system. US Patent 6,418,424 (1999) 7. Jacob, J.: Using formal specifications in the design of a human-computer interface. Commun. ACM 26(4), 259–264 (1983) 8. Johannsen, G., Levis, A., Stassen, H.G.: Theoretical problems in man-machine systems and their experimental validation. Automatica 30(2), 217–231 (1994) 9. Myers, B.: A brief history of human-computer interaction technology. Interactions 4(2), 44–54 (1984) 10. Puerta, A., Egar, J., Tu, S., Musen, M.: A multiple-method knowledge-acquisition shell for the automatic generation of knowledge-acquisition tools. Knowl. Acquis. 4(2), 171–196 (1992)

GA Based Multi-stage Transmission Network Expansion Planning A. Simo, St. Kilyeni, and C. Barbulescu(&) Power Systems Department, Power Systems Analysis and Optimization Research Center, Politehnica University Timisoara, 2, Pta. Victoriei, Timisoara, Romania {attila.simo,stefan.kilyeni, constantin.barbulescu}@upt.ro

Abstract. The paper is focusing on dynamic transmission network expansion planning (DTNEP). The DTNEP problem has been approached from the retrospective and prospective point of view. To achieve this goal, the authors are developing two software tools in Matlab environment. Power flow computing is performed using conventional methods. Optimal power flow and network expansion are performed using artificial intelligence methods. Within this field, two techniques have been tackled: particle swarm optimization (PSO) and genetic algorithms (GA). The case study refers to a real power system modeled on the Center, Northern, Eastern and Southern parts of the Romanian Power System. Keywords: Power transmission expansion
Software-tool




Power system planning




Dynamic

1 Introduction The dynamic transmission network expansion planning (DTNEP) is discussed within this paper, in retrospective and prospective manner. A set of network expansion candidates are proposed. The power flow is performed using conventional methods for all scenarios. The optimal power flow (OPF) is computed for the maximum expansion solution (including all the expansion scenarios) using particle swarm optimization (PSO) and genetic algorithms (GA). Having the optimal maximum expansion solution, the optimal expansion solution is computed also using PSO and GA. For all these purposes own software tools have been developed in Matlab environment. They are able to be linked with other well-known computer aided power system analysis software, importing the power system database. Two types of GAs are used within this paper. Binary coded GA for the expansion planning stage and real coded GA for the OPF. Dynamic programming represents an optimal solution selection methodology considering specific constraints, following a step-by-step decision process [1, 2]. Discrete dynamic programming with finite horizon is discussed within the current paper. Decisions are taken at specific time moments, following a finite number of computing steps. Currently, within the power engineering field, the heuristic and meta-heuristic methods are widely used to solve optimization problems. In case of transmission network expansion planning (TNEP), they are used to generate possible solutions, © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_5

48

A. Simo et al.

evaluation and selection of candidates, until the algorithm is not able to find a better solution. A reduced mathematical model is their main characteristic. The major drawback refers to the difficulties for tuning their parameters. A multi-stage approach of the problem under discussion is tackled in [3]. Authors are proposing a new constructive heuristic approach considering security-constrained TNEP. It is based on a local random search used to select the values of the control variables. It is applied for Ecuadorian and Chilean Power Systems. The major drawback refers to the use of a DC model. In [4] the TNEP is solved as a mixed integer nonlinear non-convex programming model. The optimization is performed based on the differential evolutionary algorithm (DEA). The use of AC load flow model is the great benefit of the paper, providing realist and accurate results. The technique is tested on Garver’s 6 bus and IEEE 24 bus test system. GA based TNEP multi-stage solving is proposed in [5]. The GA is used in conjunction with the probabilistic optimization. The objective function refers to the investment costs, absorption of private investment and system reliability. The case study refers to the IEEE 24 buses test power system. The wind generation is taken into consideration. The TNEP problem influenced by the uncertainties related to wind power generation is also investigated in [6]. In this case the Cuckoo search algorithm is used as an optimization tool. As is stipulated in [7] the use of AC mathematical model has been proposed by only very few researchers. Its complexity is augmented if it is desired to be tackled in the smart grids context. The problem is solved considered the N−1 security criterion. DEA is used and tested on Garver’s 6 bus test power system. In [8] two sources of uncertainties are analyzed in conjunction with TNEP: intermittent renewable energy generation and loads. The objective function terms are represented by the investment and operating costs. The adaptive tabu search algorithm is applied. The AC load flow mathematical model is used and interior point non-linear programing. IEEE 79 buses RTS test system is used as case study. In [9] a comparison is performed between the use of energy storage and N-1 criterion in TNEP studies. The problem is statically approached considering energy storage model, within a system with base load generators and peaking power plants. Four TNEP models are proposed for comparative analysis and N-1 network security constraints. The problem is modeled as a mixed integer linear programming problem. IEEE 24 RTS test system is used as case study. A static TNEP mathematical model is proposed in [10]. DC power flow is used. The problem is solved with and without generated power rescheduling. Artificial bee colony is used as an optimization tool for small IEEE test systems. A long-term TNEP mathematical model is proposed in [11], based on balancing investment costs and reducing consumer costs. To achieve this goal a hierarchical framework is proposed, being sensitive to different agents operating on diverse timelines. An equilibrium model is introduced combining grid operating concerns with the short-term competitive behavior of generating units. According to [12] the use of robust optimization techniques is appreciated compared with stochastic mathematical programming methods in TNEP. Several uncertainties are considered (in order to obtain more realistic results). The mathematical

GA Based Multi-stage Transmission Network Expansion Planning

49

model is represented by three-level mixed-integer optimization problem, solved using different strategies. IEEE 24 and 118 buses test systems are used. The discrete PSO approach is used in [13] to solve the TNEP problem. Non-linear mixed optimization model is used with DC power flow. Case studies are represented by IEEE 6 and 24 buses test systems and 46 buses power Brazilian system. [14] deals with DTNEP. The time evolution of the network expansion is provided considering a large number of technical and economic constraints. Results are provided only in synthesis, referring to the power system of Vietnam. Following the introduction presented within the 1st section, the 2nd one refers to the DTNEP methodology. The 3rd section deals with TNEP solving using GAs. The developed software tools are briefly presented within the 4th section. The case study is tackled within the 5th section. The 6th one is focusing on results discussing. Finally, conclusions are synthesized within the 7th section.

2 Dynamic Network Expansion Planning The DTNEP is discussed for the following time steps: 2015 year– initial stage, 2020, 2025, 2030 – intermediary stages and 2035 – final stage. The prospective approach (forward direction) and retrospective one (backward direction) has been considered. Two issues have to be solved: • consumed power forecast correlating the power generation capacity; • admissible solution domain definition – it contains the network elements’ list that are allowed to be part of the optimal solution for the final stage (2035 year). According to the prospective analysis, the starting point refers to the 2015 year. In the following, the expansion solutions are computed step-by-step for the successive years: 2020, 2025, 2030 and 2035. The provided results for 2035 year represent the final solution for the entire 20 years analyzed period. The admissible solutions’ domain has been considered to be the maximum expansion one extracting the network elements already introduced for each expansion stage. A static expansion planning solving is applied for each intermediary stage. The non-linear optimization problem is solved using evolutionary techniques: PSO and GA. According to the retrospective analysis, the starting point is represented by the maximum expansion solution, year 2035. In the following, the expansion solutions are computed step-by-step for the successive years: 2035, 2030, 2025 and 2020. The results obtained for 2035 year represents the final solution. The comments provided at the prospective analysis for the admissible solution domain definition and static expansion solving at each intermediary stage are suitable for this case too. The use of both approaches offers the advantage of comparing the intermediary solutions (2020, 2025, 2030 years) and, especially, the final one (2035 year). Mathematical model for transmission network expansion planning is presented. Two artificial intelligence solving techniques have been tackled: PSO and GA. The optimization problem has a multi-criteria character. The following components are included within the global objective function:

50

A. Simo et al.

The optimization problem has a multi-criteria character. The following components are included within the global objective function: • power system operating costs: OBF1 ¼

X

Ci ðPg i Þ þ

X

TPij ðSij  S ij Þ ¼ Minim

ð1Þ

ij2R

i2G

where: G – set of PV buses; R – set of network elements; Ci(Pgi) – quadratic cost characteristics of generators – Ci ðPgi Þ ¼ ai
P2gi þ bi
Pgi þ ci ; TPij – penalty cost of the apparent power upper limit exceed trough the ij network element; S ij is defined  max Sij if Sij  Sij as S ; Sij – apparent power flow for ij network element; ij ¼ Smax if Sij [ Smax ij ij max Sij – Sij thermal limit; • investment equivalent yearly cost related to new power transmission capacities (overhead lines, autotransformers) – OBF2; • safety operation, quantified based on risk factor computing: n‘ P

OBF3 ¼ r % ¼

n‘ P

qk
r k

k¼1 n‘ P

¼

k¼1

n  o   qk
Pkr  Skij  [ Smax ; ij 2 R ij n‘ P

qk

k¼1


100

ð2Þ

qk

k¼1

where: qk – k overhead line (OHL) disconnection probability; nl – overhead lines’ number that are selected for contingencies; Skij – Sij in case of k element disconnection; rk – congestion probability when k OHL is disconnected. • total available transmission capacity. OBF4 ¼ TATC ¼

X



 ij 2 Lmax Sij \S ij

  Sij  Smax Þ  ij

ð3Þ

The function that has to be minimized is represented by the sum of the suitable weighted four components previously discussed (wi – weighting coefficients): OBF ¼ w1
OBF1 þ w2
OBF2 þ w3
OBF3 þ w4
OBF4

ð4Þ

3 Genetic Algorithm (GA) Based TNEP The population represents a set of possible solutions. Each chromosome contains binary digits (0, 1), representing the state for the network expansion candidates. Thus, for this stage we are dealing with binary coded GA.

GA Based Multi-stage Transmission Network Expansion Planning

51

The chromosome has d length, being able to be written as: xi ¼ fxi1 ; xi2 ; . . .; xid g;

i ¼ 1; 2;


; nc

ð5Þ

Each chromosome is evaluated based on the OBF objective function. The computing process finishes if the solution is not able to be improved. The algorithm stages are the following ones: (a) chromosomes forming the population are randomly initialized with 0 and 1: x0i ¼ fx0i;1 ; x0i;2 ; . . .; x0i;d g;

i ¼ 1; 2;


; np

ð6Þ

(b) GA based OPF is computed for the configuration coded by each chromosome; (c) initial population is evaluated based on OBF value. The best chromosome is saved in x0elit : f ðx0elit Þ ¼ minfOBFðx0i Þg; i ¼ 1; 2; . . .; nc

ð7Þ

(d) for a specific t computing step (t = 0, 1, 2,…) the chromosomes forming the population subjected to recombination are selected; (e) npr ¼ v
nc =2 chromosome pairs that are subjected to crossover are formed and chromosome pairs that are going to be copied unaltered; (f) offspring are formed starting from the npr pairs subjected to crossover (in one point or in several points); (g) number of chromosome genes subjected to mutation is computed; (h) 1st chromosome belonging to the population obtained at previous step is replaced with the best of the old population: x1t þ 1 ¼ xtelit

ð8Þ

(i) OPF is computed for the configurations coded by each chromosome. Current population is evaluated based on OBF value. New xelit value is computed: þ1 f ðxtelit Þ ¼ minfFOBðxti þ 1 Þg;

i ¼ 1; 2; . . .; nc

ð9Þ

(j) if the OBF value is not able to be improved for a given number of steps, the computing process finishes. The operating condition corresponding to the last xelit value represents the optimal one. Contrary, computing step is increased with 1 and the algorithm is repeated starting with point c).

4 Software Tool Two software tools have been developed by the authors in Matlab environment based on PSO and GA approaches: Power OptPowerplanPSO and PowerOptPowerPlanGA. Each one has two modules linked through a graphical user interface:

52

A. Simo et al.

• 1st module – used for OPF (also is able to be used as a stand-alone module); • 2nd module – used for dynamic transmission network expansion planning. PowerOptPowerplanGA’s main window for the OPF module is presented in Fig. 1. Once the power system database has been loaded, the user is able to select the optimization type he desires by selecting the control variables. The lower part of the window allows the user to set the GA parameters: maximum computing steps, capping iterations, population size etc. The main window for the software tool 2nd module (expansion planning module) is presented in Fig. 2. The user has to set the parameters for the GA algorithm: selection type, recombination type etc.

Fig. 1. PowerOptPowerplanGA – OPF module main window

Fig. 2. PowerOptPowerplanGA – dynamic network expansion window

Computing-candidates tab allows the user to specify if for each of the power system configurations corresponding to different expansion candidates, OPF is computed or only power flow. Button labeled Number of expansion stage [years] allows the user to specify the desired number of expansion stages. The entire expansion planning time horizon is going to be divided into the specified number of years. The results are provided in several ways. They are exported in different file formats (according to the user’s desire), view on the display or graphically displayed. The 2nd software tool based on PSO is named PowerOptPowerPlanPSO. This one is not discussed within the paper.

GA Based Multi-stage Transmission Network Expansion Planning

53

5 Case Study The power system used as case study has been modeled based on the Center & Northern & Eastern & Southern parts of the Romanian Power System. The other parts (Western, South-Western and North-Western) have been eliminated. The one-line diagram is presented in Fig. 3 having the following characteristics: • number of buses – 110 (30 PV buses and 80 PQ buses); • number of network elements – 149 divided as: 105 overhead lines of 110, 220 and 400 kV, 44 (auto)transformers; • number of generating units – 30, 23 being real ones and 7 equivalent ones.

Fig. 3. Case study one-line diagram

54

A. Simo et al.

6 Results and Discussions The dynamic transmission network expansion planning is performed (retrospective and prospective approach). This process is step-by-step discussed within the following subsections. 6.1

Base Case – Year 2015

Base case has been computed using conventional methods. Bus voltages are ranging between 0.95 and 1.10 p.u. (110 kV, 220 kV), respectively 1 and 1.05 p.u. (400 kV). PV buses terminal voltage limits are set between 0.95 and 1.15 p.u. The total consumed power is Pc = 4172.8 MW, the real generated power Pg = 4289.7 MW and real power losses ΔP = 116.9 MW. 6.2

Maximum Expansion Solution – Year 2035

The transmission network expansion planning is discussed for a 20 years, based on the last year consumed power forecast. The 2035 year forecasted peak-evening-winter operating condition is characterized by total real consumed power of 7473 MW (in comparison with 4172.8 MW for the base case), respectively reactive power 2945 MVAr (compared with 1999.9 MVAr). For load fulfilment the following generating units have been introduced: • 330 MW units at 28061 (2), 28904 (2), 28021 (3) 28022 (2) buses; • 220 MW units at 28077 (2) bus; • wind farms injecting power in 28017 (600 MW), 28974 (200 MW), 28019 (160 MW) and 28080 (100 MW) buses. A number of 23 new transmission network elements (20 OHLs and 3 autotransformers) have been introduced (considered as candidates within the expansion list). Thus, the expansion scenario is the following one: • new 400 kV buses at 28040, 28062 and 28908; • 400/220 kV 400 MVA (auto)transformers in substations 28061–28062 (2 units) and 28907–28908; • new 400 kV OHL 28037–28040, 28040–28082, 28031–28908, 28908 – 28016; • 220 kV OHL upgraded to 400 kV: 28082–28950, 28950–28025, 28905–28906, 28906–28908; • 400 kV OHL additional circuit at 28016–28973, 28022–28024; • 220 kV OHL additional circuit at 28045–28061, 28061–28057, 28057–28055, 28087–28086 (2), 28086–28085, 28085–28084, 28084–28083, 28083–28078, 28078–28024. Using the software tool PowerOptPlanGA the optimal power flow – OPF has been computed for the maximum expansion solution. Real power losses are equal to ΔP = 191.8 MW, compared to 225.1 MW for the maximum expansion solution base case. Thus, about 15% decreasing has been recorded.

GA Based Multi-stage Transmission Network Expansion Planning

55

The GA algorithm evolution for OPF is presented in Fig. 4. The OBF individual best values are presented using blue colour for each computing step. The OBF average value for the entire population is represented using green colour. The OBF value, corresponding to the worst solution, is represented using red colour. An accentuated decrease of the OBF value is recorded during the 1st 50 computing steps. The solution is slightly improved for the next computing steps. Although we are dealing with a large scale power system, the number of the computing steps required for the solution computing is reduced. This fact proves the correctness of the GA parameters’ tuning.

Fig. 4. GA algorithm evolution for OPF computing

6.3

Optimal Expansion Solution

• Retrospective approach Before TNEP the load forecast for 20 years period (2015–2035) has been performed. The results are presented in Table 1. Table 1. Consumed power forecasting 2015 2020 2025 2030 2035

Pc total [MW] 4172.8 4827.2 5584.2 6459.9 7473.0

Qc total [MVAr] 1999.9 2203.1 2426.9 2673.4 2945.0

56

A. Simo et al.

The starting point is represented by the maximum expansion solution, year 2035 and the expansion solutions for all the stages: 2035, 2030, 2025 and 2020. The solution admissible domain for each dynamic expansion stage has been defined based on the previous results. For DTNEP the PowerOptPowerPlanGA software tool has been used. The following elements have been found for the optimal expansion solution: • new 400 kV buses at 28040, 28062 and 28908; • 400/220 kV 400 MVA (auto)transformers in substations 28061–28062 (2 units) and 28907–28908; • new 400 kV OHL 28037–28040, 28040–28082 (alternative solution – double circuit 220 kV OHL), 28031–28908, 28908–28016; • 220 kV OHL upgraded to 400 kV: 28082–28950, 28950–28025; • 220 kV OHL additional circuit at 28061–28057, 28057–28055, 28087–28086 (2), 28086–28085, 28083–28078, 28078–28024 buses. The optimal expansion solution is characterized by 124 buses (41 PV buses, 83 PQ buses) and 176 network elements (118 OHL, 58 transformers, autotransformers). The relative OBF value (Fig. 5) has been computed being the ratio between the current expansion solution OBF and the one corresponding to the maximum expansion solution. The OBF value permanently improves during the algorithm evolution. This fact is explained due to the power system scale and increased number of transmission network expansion candidates. For the 1st three computing steps 17.5% OBF decreasing value is recorded and additionally 1.2% for the following ones. The dynamic retrospective transmission network expansion planning results are synthesized in Table 2.

Fig. 5. GA based OBF evolution

They are several quasi-optimal solutions, having close OBF values. As an alternative solution, the 28037–28040, 28040–28082 400 kV OHL’s are replaced by d.c. 220 kV OHL’s.

GA Based Multi-stage Transmission Network Expansion Planning

57

Table 2. Retrospective analysis results 2020 3 from 20

2025 7 from 20

2030 12 from 20

OHL – – – – – – – – 400 – – 400 – – 400 – 400 kV 28022–28024 400 400 kV 28082–28950 400 kV 28082–28950 400 400 kV 28082–28950 400 kV 28082–28950 400 220 kV 28057–28055 220 kV 28061–28057 220 – 220 kV 28057–28055 220 – 220 kV 28087–28086 220 – – – – 220 kV 28086–28085 220 – – 220 – – 220

kV kV kV kV kV kV kV kV kV

28031–28908 28908–28016 28016–28973 28022–28024 28082–28950 28082–28950 28061–28057 28057–28055 28087–28086

kV 28086–28085 kV 28083–28078 kV 28078–28024

2035 15 from 20 400 400 400 400 400 400 400 400 220 220 220 220 220 220 220

kV kV kV kV kV kV kV kV kV kV kV kV kV kV kV

28037–28040 28040–28082 28031–28908 28908–28016 28016–28973 28022–28024 28082–28950 28950–28025 28061–28057 28057–28055 28087–28086 28087–28086 28086–28085 28083–28078 28078–28024

• Prospective approach The starting point is represented by the initial situation corresponding to the 2015 year. Expansion solutions for each future stage are computed step-by-step (2020, 2025, 2030, 2035 years). Results obtained for 2035 year are representing the final solutions for the 20 years’ analyzed period. The admissible solutions’ domain has been considered to be the one defined by the maximum expansion solution, excluding the network elements already introduced at each stage of the prospective dynamic expansion. The dynamic prospective transmission network expansion planning results are synthesized in Table 3. 6.4

Comments on the Expansion Solution

Comparing the results gathered from both approaches the following conclusions are highlighted: • 2020 year solution is different for the two approaches, but it has very close OBF value. For the prospective approach the 28057–28055 and 28087–28086 220 kV OHLs have been considered as double circuit, resulting 5 new network elements (instead of 3 in the retrospective one); • a similar comment is suitable for 2025 year solution – the total number of new network elements is 8 in the prospective approach (7 in the retrospective one). The difference refers to the 28016–28973 220 kV OHL which has been considered already as double circuit (in the retrospective solution only in 2030);

58

A. Simo et al. Table 3. Prospective analysis results 2020 5 from 20

OHL – – – – – – 400 400 220 220 220 – – – –

kV kV kV kV kV

28082–28950 28082–28950 28057–28055 28057–28055 28087–28086

2025 8 from 20

2030 13 from 20

2035 15 from 20

– – – – 400 400 400 400 220 220 220 – 220 – –

– – 400 400 400 400 400 400 220 220 220 220 220 220 220

400 400 400 400 400 400 400 400 220 220 220 220 220 220 220

kV kV kV kV kV kV kV

28016v28973 28022–28024 28082–28950 28082–28950 28061–28057 28057–28055 28087–28086

kV 28086–28085

kV kV kV kV kV kV kV kV kV kV kV kV kV

28031–28908 28908–28016 28016–28973 28022–28024 28082–28950 28082–28950 28061–28057 28057–28055 28087–28086 28087–28086 28086–28085 28083–28078 28078–28024

kV kV kV kV kV kV kV kV kV kV kV kV kV kV kV

28037–28040 28040–28082 28031–28908 28908–28016 28016–28973 28022–28024 28082–28950 28950–28025 28061–28057 28057–28055 28087–28086 28087–28086 28086–28085 28083–28078 28078–28024

• for 2030 year solution the total number of new network elements is 13 in the prospective approach (instead of 12 in the retrospective one). The difference refers to the 28087–28086 220 kV OHL which has been considered already as double circuit (in the retrospective solution in 2035); • the final expansion solution is the same for the both approaches, the intermediary stages being slightly different.

7 Conclusions The developed software tools are able to be used in case of large scale, complex transmission networks. They behave as hybrid software tool, the PSO and GA techniques being used for the OPF and network expansion stages. PSO and GA have been adapted to the power engineering field. The prospective and retrospective dynamic expansion approaches are providing identical solutions (for the last year). For both cases other expansion solutions, additional to the optimal one, have been proposed. This information is helping the network planner to adopt the best planning scenario. However, some difference may appear during the intermediary stages (years). In case of the prospective approach, there are situations when the process is started from an initial solution less or more different from the retrospective approach. But, the final results are the same. There are also situations when the solutions for the 1st and final expansion stage (year) are the same, but intermediary ones (2025, 2030 years) are different.

GA Based Multi-stage Transmission Network Expansion Planning

59

References 1. Bellman, R.E., Dreyfus, E.S.: Applied Dynamic Programming. Princeton University Corporation, Princeton (1962) 2. Sniedovich, M.: Dynamic Programming. Marcel Dekker, Inc., New York City (1992) 3. Hinojosa, V.H., Galleguillos, N., Nuques, B.: A simulated rebounding algorithm applied to the multi-stage security-constrained transmission expansion planning in power systems. Electr. Power Energy Syst. 47, 168–180 (2013). Elsevier 4. Alhamrouni, A., Khairuddin, A., Khorasani, F., Salem, M.: Transmission expansion planning using AC-based differential evolution algorithm. IET Gener. Transm. Distrib. 10, 1637–1644 (2014) 5. Arabali, A., Ghofrani, M., Amoli, E.M., Fadali, M.S., Aghtaie, M.M.: A multi-objective transmission expansion planning framework in deregulated power systems with wind generation. IEEE Trans. Power Syst. 6, 1–8 (2014) 6. Taheri, S., Seyed-Shenava, S.J., Modiri-Delshad, M.: Transmission network expansion planning under wind farm uncertainties using Cuckoo search algorithm. In: Proceedings of the 3rd IET International Conference on Clean Energy and Technology (CEAT), 24–26 November, pp. 1–6 (2014) 7. Torres, S.P., de Araujo, R.A., Castro, C.A., Pissolato, J.: Security constrained transmission expansion planning for smart transmission grids based on the AC network model. In: IEEE PES Transmission and Distribution Conference and Exposition – Latin America (PES T&D-LA), 10–13 September, pp. 1–6 (2014) 8. Chatthaworn, R., Chaitusaney, S.: Improving method of robust transmission network expansion planning considering intermittent renewable energy generation and loads. IET Gener. Transm. Distrib. 13, 1621–1627 (2015) 9. Obio, E.B., Mutale, J.: A comparative analysis of energy storage and N−1 network security in transmission expansion planning. In: Proceedings of the IEEE 50th International Universities Power Engineering Conference (UPEC), 1–4 September, pp. 1–6 (2015) 10. Rathore, C., Roy, R.: Gbest-guided artificial bee colony algorithm based static transmission network expansion planning. In: Proceedings of the IEEE International Conference on Electrical, Computer and Communication Technologies (ICECCT), 5–7 March, pp. 1–6 (2015) 11. Tang, L., Ferris, M.C.: A hierarchical framework for long-term power planning models. IEEE Trans. Power Syst. 1, 46–56 (2015) 12. Mínguez, R., García-Bertrand, R.: Robust transmission network expansion planning in energy systems: improving computational performance. Eur. J. Oper. Res. 248, 21–32 (2015) 13. Barreto, W.E., Torres, S.P., Castro, C.A.: Study of particle swarm optimization variations applied to transmission expansion planning. In: Proceedings of the IEEE Powertech Conference, 16–20 June, pp. 1–6 (2013) 14. Le, A.D., Nguzen, M.H., Eghbal, M., Nguzen, D.H.: Transmission expansion planning in electricity market: the case in Vietnam. In: Proceedings of the IEEE Australasian Universities Power Engineering Conference (AUPEC), Hobart, Australia, 29 September– 03 December, pp. 1–6 (2013)

Epidemic Algorithm Based Optimal Power Flow in Electric Grids K. Muniyasamy1 , Seshadhri Srinivasan2(B) , S. Parthasarathy3 , B. Subathra4 , and Simona Dzitac5 1

Department of Electrical and Electronics Engineering, Kalasalingam University, Virudhunagar, India [email protected] 2 International Research Center, Kalasalingam University, Virudhunagar, India [email protected] 3 Department of Electrical and Electronics Engineering, KLN College of Engineering, Sivagangai, India sarathy [email protected] 4 Department of Instrumentation and Control Engineering, Kalasalingam University, Virudhunagar, India [email protected] 5 University of Oradea, Oradea, Romania [email protected]

Abstract. Time-triggering and distributed nature of the grid are emerging as the major challenge in managing energy in distribution grids. This investigation presents an event triggered distributed optimal power flow (OPF) algorithm for energy grids. To generate the event triggers, we use the epidemic algorithm. The buses are classified into three: infected, susceptible, and dead. The network works in two modes: normal and optimization mode. In the normal mode, only event detection happens and when there are no event triggers, the system is said to be in normal mode. In optimization mode, event triggers that can be a change in generation or demand beyond a threshold value that necessitates the reoptimization of the network, the optimization mode begins. In this mode, the infected node which is infected by change in bus variable intimates it to the energy management application. The energy management application on sensing this change, will initiate the graph grammars which are a set of rules to change the bus nature by detecting the effect of the change on the particular bus. The network is re-optimized using a DC OPF formulation as it is convex and can be solved using simple matrix inversion on the stationary conditions. As a result, the solution of DCOPF problem becomes that of solving a system of linear equations of the form Ax = b, which is solved using Krylov’s method or the Arnoldi algorithm in a distributed fashion. Each node solves the problem of its one-hop neighbours in parallel and this leads to a distributed implementation resulting significant reduction in complexity. The propossed approach is illustrated on a simple 3 bus network. c Springer International Publishing AG 2018  V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8 6

Epidemic Algorithm Based Optimal Power Flow in Electric Grids

61

Nomenclature i, j θ B xij Dt Pt PGg g t C(PGg ) T λ G Km

1

indices of bus load angles admittance matrix reactance of line ij Vector of active demand at time t Vector of active power generation at time t Power generated by the generator g indices for generator time indices Generation cost of the generator g in dollars Set of all time Lagrange Multiplier set of all generators Krylov’s Space

Introduction

In many energy management systems (EMS), time-triggering is used for optimization (see, [1–5] and references therein). This leads to unnecessary computations as the grid re-optimizes even without change in grid conditions. The complexity is compounded by the distributed nature of the grid. Considering the fluctuating nature of the demand and generation, and their distribution within the grid, event-triggering is more suitable as thresholds of change are detected and re-optimization of the EMS application is initiated. Moreover, some of these changes can be localized using a distributed optimization algorithm. This significantly reduces the complexity. Therefore, energy management applications that are event triggered and perform distributed optimization are very much required in modern energy grids integrated with renewable energy sources (RES) and energy storage systems (ESS) [6]. Optimal Power Flow (OPF) is a widely used tool for energy management and planning. The OPF problem computes the optimal generator dispatches considering the network parameters to optimize either the cost, line-losses, demand response, and many other objectives (see, [7–12]). There are various solution methods for OPF such as quadratic programming [13], semi-definite programming [14], sequential quadratic programming [15], linear programming [16], and nonlinear programming [22,27] to name a few. These approaches use a static optimization approach and do not consider time-variations of the demand or fluctuating generation and are hence not suitable for modern grids. Dynamic OPF overcomes this problem as it solves a problem considering time as a running parameter [17]. However, still the dynamic OPF cannot deal with fluctuations in demand and generation. To overcome this receding

62

K. Muniyasamy et al.

horizon approach or model predictive control for OPF problem has been studied. In [22,28], the ACOPF problem using forecast on renewable generation and demand has been solved using the receding horizon approach. The major drawback with these approaches is the time-triggering. The grid is re-optimized even when there is no significant changes in grid conditions or optimized only after a certain change has occurred which is not useful aspect of the EMS. A modified scenario could be to initiate the optimization whenever there are significant changes in the grid variables such as demand or generation. This will reduce frequent optimization and computation overhead resulting thereof. More recently, distributed approaches for solving the OPF problem based on Gradient descent [19], decomposition methods [21], alternating directional method of multipliers [18], and others have been studied. In principle these methods consider two types of network models: ACOPF and DCOPF. While ACOPF provides a comprehensive study on both transient and steady state behaviour including the voltage magnitudes, they are non-convex and non-linear requiring significant computations for optimizing the grid. On the other hand, in many scenarios DCOPF which are an approximation of ACOPF problem behave reasonably well and are widely used in practice [20]. The main advantage of DCOPF is that it leads to a convex problem and its solution is relatively simple and unique compared to ACOPF. Moreover, as just a matrix inversion on the Karush-Kuhn Tucker (KKT) conditions or the stationary conditions of the DCOPF problem can provide the optimal power dispatches, they are widely preferred. This investigation addresses the two challenges discussed above timetriggering and distributed nature of the grid. It uses an event-triggered optimization and solves the DCOPF problem in a distributed manner. To generate the event triggers, it uses the epidemic programming widely studied for spreading information in distributed systems [23]. Then it solves the DCOPF problem in a distributed manner by formulating it as a set of linear equations and uses the Krylov’s method (i.e. Arnoldi’s algorithm) as the stationary conditions of the DCOPF problem lead to a positive definite matrix. The solution can be obtained fast using the Krylov’s method. The main contributions of the investigation are: • An epidemic algorithm based event triggering mechanism. • A self-assembly procedure using rules. • Distributed OPF implementation using Krylov’s method. The rest of the paper is organized as follows. Section 2 presents the problem formulation. The concept of autonomous gossiping is presented in Sect. 3. The Krylov’s method for performing distributed optimization is presented in Sect. 4. Section 5 presents a 3-bus case study.

Epidemic Algorithm Based Optimal Power Flow in Electric Grids

2

63

Problem Formulation

Consider the multi-period DCOPF problem widely studied in literature:   C (Pg ) t∈T pg ∈G

Pg

s.t.

(1) B.θt = Pt − Dt , Pgmin

≤ Pg, t ≤

∀t ∈ T

Pgmax

The line-flow constraints and the ramp-up/down constraints of the generators are not considered in the formulation due to simplicity. One can verify that the variations can happen in the optimization problem only due to changes in real-power generations which may be due to renewable generation or changes in demand. The problem is usually solved periodically based using time as a reference parameter. However, performing time-triggered optimization requires the problem be solved repeatedly even when there is no update in the generation. This is generally not preferred. A more realistic scenario can be the event triggering i.e., on sensing there is a change in grid condition re-optimize the generator settings for reducing the operating cost. But again the problem has to be solved in a centralized fashion which increases the computation complexity. Significant reduction can be obtained, if the problem is solved in a distributed fashion. Moreover, since the DCOPF problem is convex, the KKT conditions can provide the unique solution to the optimization problem. Further, the very nature of the network architecture i.e. the bus reactance from bus i to j, xij is equal to in magnitude to xji . Therefore, elements of the matrix that models the stationary conditions is symmetric. The problem is to study the design of a distributed DCOPF implementation that uses event triggering for optimizing the network operating cost. The formulation leads to the following pertinent questions: 1. How to generate the event triggers? 2. How to distribute the computation? 3. Can convergence and robustness of the distributed computation be guaranteed? The answers to the above questions will be considered in the rest of the paper.

3

Autonomous Gossiping

Autonomous gossiping algorithms such as the epidemic algorithms are emerging as a method for propagating information in distributed systems [24]. They have been applied to wide class of problems like replication of database systems [25], wireless sensor routing, and more recently also in electric networks [26]. They

64

K. Muniyasamy et al.

are seen as a robust and scalable means to disseminate information reliably on a distributed network the same way an epidemic will propagate through a set of individuals. The use of gossip-like distributed optimization algorithm for reactive power flow control was investigated by Bolognani and Zampieri in [26]. The investigation considered its use in the reactive power flow control problem of the microgrid. This investigation considers the use of the autonomous gossip algorithms for distributed optimal power flow (DOPF). Each bus is considered one agent in the algorithm. The agents with the generators have a cost function that models the generation cost that is quadratic. Hence, for each generation node, the cost is given by C(PG g) = ag × (PGg )2 + bg × (PGg ) + cg

∀g ∈ G

where ag , bg and cg are constants provided by the GenCo. We assume that an event occurs whenever there is a change in the generator or demand settings and the general optimization problem is less sensitive to changes below the threshold i in each agent. The threshold value can be determined from sensitivity analysis or by simple numerical iterations. In some instances, the values can be updated online as well. The algorithm works in two modes: normal and optimization. We assume that the optimal generator settings for the normal mode are obtained using a distributed OPF formulation and that the grid is functioning within thresholds. The normal mode monitors for the threshold  being breached in some node and sends triggers for re-optimization. As soon as the triggers are received, the algorithm enters the optimization mode, wherein the nodes are classified into three categories: infected, susceptible, and dead. The infected node is the one that has exceeded the threshold, whereas the susceptible can contract the infection, and the dead-node is not affected directly by the infected node, or is not affected at all. Three types of actions are defined: push, pull, and push-pull. In the push action, the infected node tries to send information to its one-hop neighbours regarding the event. The susceptible nodes try to pull the information from the infected node on the magnitude of the violation. Then they inform the dead node of the violation and also pull an information from them about their current status. In order to promote autonomous gossiping, we define the rules in Table 1. The above table is used to create autonomous organization of the agents and they form a self-assembly. The agents status gets changed during each iteration of the algorithm until the threshold reaches below i in all the infected buses. The algorithm starts with the infected node pushing its information to its onehop neighbours during the first iteration. The susceptible nodes receive this information and start re-optimizing triggers to their peers. The common variable is the load angle θi that models the difference between the load balance and also the line-limits. On sensing a variation in theta from the optimal value, the network is broken at the very node that experiences it, and various zones are formed. The coupling

Epidemic Algorithm Based Optimal Power Flow in Electric Grids

65

between the different zones is the variable θi that is replicated. The self-assembly step will be used to send event trigger and the algorithm continues its execution until the convergence to optimal value. Table 1. Self-organizing rules for the nodes in DOPF

4 4.1

Node 1

Node 2

Infected

Susceptible Push

Action

Node 1: Infected, Node 2: Infected

Outcome

Infected

Susceptible Pull

Node 1: Infected, Node 2: Susceptible

Infected

Susceptible Push-pull Node 1: Infected, Node 2: Infected

Susceptible Infected

Pull

Node 1: Infected, Node 2: Infected

Susceptible Infected

Push

Node 1: Susceptible, Node 2: Infected

Susceptible Infected

Push-pull Node 1: Infected, Node 2: Infected

Infected

Dead

Push

Node 1: Infected, Node 2: Susceptible

Infected

Dead

Pull

Node 1: Infected, Node 2: Dead

Infected

Susceptible Push-pull Node 1: Infected, Node 2: Susceptible

Susceptible Dead

Pull

Node 1: Susceptible, Node 2: Dead

Susceptible Dead

Push

Node 1: Susceptible, Node 2: Susceptible

Susceptible Dead

Push-pull Node 1: Infected, Node 2: Susceptible

Krylov’s Method Based Distributed Optimization Langarangian of the DCOPF Problem

The Lagrangian of the DCOPF problem is given by L =

N 

(ai × Pg2 + bi × Pg + ci ) + λi ((Bθ − Pi + Di ) + μ(θ1 − 0))

(2)

i=1

where θ1 is the reference bus. The optimal value of the DCOPF problem can be obtained from the KKT conditions: • First order necessary conditions

• Primal feasibility

∂L =0 ∂Pg

(3)

∂L =0 ∂θ

(4)

∂L =0 ∂λi

(5)

66

K. Muniyasamy et al.

• Dual feasibility

λi ≥ 0

(6)

λi (Bθ − Pi + Di ) = 0

(7)

• Complimentary Slackness

The above conditions lead to a set of linear equations in the unknowns X = [Pg , θ, λ]. The problem can hence be represented as: M ×X =b

(8)

The above solution has 3 × n linear equations that need to be solved in a distributed fashion. The objective of this investigation is to have a distributed approach, which means the computation above has to be distributed between the different nodes. To form the distributed updates, the infected and its one hop neighbours are considered, a reduced order linear equations of the one-hop neighbours is solved using the Krylov’s method. The idea is to construct a reduced space Km with a dimension m 0 are weighing matrices of appropriate dimensions, and QN ≥ 0 is the terminal weighing matrix. Proof: Define the cost-to-go function at time k as, Jk (xk ) =

min

uk ,...uN −1

N −1 

((xχ − rχ )T Q(xχ − rχ ) + uTχ Ruχ )

χ=k

+ (xN − rN )T QN (xN − rN ) subject to relations (Eq.(5))

(12)

Note that Jk (xk ) gives the minimum cost-to-go starting from the state xk at time instant k. As Jk (xk ) is quadratic, the cost-to-go has the form (Eq. (13)) Jk (xk ) = xTk Pk xk + 2qk xk + rkT Qrk

(13)

where qk = −QT rk , and Pk = PkT ≥ 0 that can be found recursively using backward iteration algorithm by taking PN = QN .

86

S. Srinivasan et al.

Assuming that Jk+1 has the given form, from dynamic programming principle, suppose we know Jk+1 (xk+1 ) and look for optimal input uk the Bellman equation is given by (Eq. (14)) Jk (xk ) = min(xk − rk )T Q(xk − rk ) + uTk Ruk u

+ Jk+1 (φxk + Γ0 uk + Γ1 uk−1 ) = (xk − rk )T Q(xk − rk ) + min(uTk R uk + (φxk + Γ0 uk + Γ1 uk−1 )T

(14)

u

Pk+1 (φxk + Γ0 uk + Γ1 uk−1 ) T (φxk + Γ0 uk + Γ1 uk−1 ) + 2qk+1

From the necessary conditions of optimality, we obtain Ruk + Γ0T Pk+1 [φxk + Γ0 uk + Γ1 uk−1 ] + Γ0T qk+1 = 0

(15)

Little manipulation of Eq. (15) leads to Eq. (9) with the adaptive gains given by Eq. (10).  Obviously, the computation of adaptive controller gains requires the knowledge of Pk+1 and the estimate of time-delay τk to compute Γ0 (τk ) and Γ1 (τk ). We compute Pk+1 recursively working backward from k = N by letting PN = QN , where QN is the terminal weighing matrix. The estimate of τk is obtained as described in Sect. 3. The Gains are computed from Eq. (10). Having obtained the results required for computing the gains, we now proceed to illustrate the controller by combining experiments with simulation.

5

Results and Discussions

This section presents the results obtained from experiments and simulations to illustrate the performance of the proposed adaptive controller. 5.1

Delay Estimation from Experiments

To estimate the time-varying delays, experimentation was conducted on Modbus over TCP/IP by connecting several controllers, and then loading the network (by increasing the number of inputs/outputs and number of rungs in PLC ladder diagram). Further, the length of the communication channel was varied by connecting controllers such as ABB AC 500 at different locations. The prototype of the experiment is shown in Fig. 5 wherein there is a Master controller (also called the Modbus Master), and Slave controllers. The delays for variations of the parameters during experimentation is recorded both for training and validation phases of MRAN. During the network training phase the weights were adjusted to match the samples, whereas in the validation phase the delay estimates were compared with the estimates

Adaptive Controller for Networked Control Systems Subjected

87

Fig. 5. Experiment pilot for delay estimation

generated by MRAN. The first 400 samples of delay data were used in learning, and the next 100 samples of data were used in validation of the estimates. After validation, in the test phase MRAN can be used to generate the delay estimates for designing the adaptive controller on-line. The number of hidden neurons in MRAN, delay samples from the network, estimated delays, and the error (difference between the actual and estimated network delay) are shown in Fig. 6. It was found that MRAN can estimate the delay with a maximum error of 2.2 ms, and an average error of less than 1 ms over 10 samples. 30

25

no. of neurons in hidden layer Actual and Estimated Output Error of Actual and Estimated Output

20

No of Nueurons Actual Output Estimated Output error

15 Training Data

Testing Data

10

5

0

−5

0

50

100

150

200

250 300 Sample points

350

400

450

500

Fig. 6. Combined plot: number of Hidden neuron, actual, estimated output and error

88

S. Srinivasan et al. 10 h=0.01 s delay=0.001 s

8

h=0.01 s delay=0.005s

x1(t)

6 4 2 0 −2

0

5

10

15

20

25 30 time in Secs

35

40

45

50

0

5

10

15

20

25 30 time in Secs

35

40

45

50

5 0

u

k

−5 −10 −15 −20

Fig. 7. Response of the system (Eq. (16)) to the proposed adaptive controller and fixed delays

5.2

Example 1

To illustrate MRAN-based adaptive controller, we consider the scalar example x(t) ˙ = 2x(t) + u(t)

(16)

with rk = 1 and initial condition x0 = 2 as in [12]. The poor performance of the static state feedback controller for this example has been illustrated by the authors of [14]. Our investigation uses a sampling time of 10 ms, and the estimated delay varies from 4 ms to 10 ms. Figures 7 and 8 show the response along with the control input uk with delays being changed at t = 25 s and delay varying at every 10 s using delay estimates from MRAN, respectively. This result illustrates that MRAN based controller adapts the gain to track the controller input even in the case of fast variations in delay. 5.3

Example 2

The investigation considers the double integrator system     01 0 x(t) ˙ = x(t) + u(t) 00 1   y(t) = 1 0 x(t)

(17)

to illustrate the effectiveness of the proposed adaptive regulator. The continuoustime dynamics (Eq. (17)) is discretized with sampling period h under the assumption that τk ≤ h, resulting in Eq. (18)

Adaptive Controller for Networked Control Systems Subjected

89

10

x1(t)

5 0 −5

0

5

10

15

20

25 30 time in secs

35

40

45

50

0

5

10

15

20

25 30 time in secs

35

40

45

50

0

5

10

15

20

25 30 time in secs

35

40

45

50

9

τk

8 7 6 5

10

u

k

0 −10 −20

Fig. 8. Response of system (Eq. (16)) to the proposed adaptive controller and variable delays, estimated by MRAN

   (h−τk )2 1h τ (h − 2 = xk + uk + k 01 τk h − τk 

xk+1



τk 2 )

uk−1

(18)

The sampled-data representation of linear NCSs with total delay (at each sampling instant k) less than the sampling time, i.e. τk ≤ h, can be obtained using MATLAB routine NCSd developed by the authors of the paper, see [33]. Figure 9 shows the response of the system (Eq. 18) under the MRAN-based adaptive controller ((Eq. 5) using the constant reference signal rk = [0 0]T for each k. The Figure additionally shows the control values uk and the delay estimates from MRAN. The result shows the regulatory performance of the controller in the presence of delays. Variations in the gains Lx , Lr , and Lu with time-varying delays τk are shown in Fig. 10. The gains are adjusted according to channel delays τk to track the reference state, and these variations are reflected mostly in the gain Lu . The states of the double integrator system in Eq. (18) with the adaptive controller (Eq. (5)) and desired states rk = [1 0]T are shown in Fig. 11. This response illustrates the tracking performance of the adaptive controller. Variations in gain Lx , Lr , and Lu with time-varying delays τk are shown in Fig. 12.

90

S. Srinivasan et al.

0.02 x1(t)

0.01 0 −0.01

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

70

80

90

100

x (t) 2

0.1 0 −0.1

k

0.5

u

0 −0.5

τ

k

8 6 4

Fig. 9. States of the system and the control input with time-varying delay

x

L (k)

0 −0.5 −1 −1.5

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50 60 time in secs

70

80

90

100

Lr(k)

0 −1 −2

Lu(k)

0 −0.01 −0.02

τk

10 5 0

Fig. 10. Variation in gains Lx , Lr , and Lu with delays for regulatory performance

Adaptive Controller for Networked Control Systems Subjected

91

1.5

1

x (t)

1 0.5 0

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50 60 time in secs

70

80

90

100

x (t) 2

0.5 0 −0.5 2 uk

1 0 −1

τk

8 6 4

Fig. 11. States of the system and the control input for tracking the reference with time-varying delays (x0 = [0 0]T )

−0.5

x

L (k)

0

−1 −1.5

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50

60

70

80

90

100

0

10

20

30

40

50 60 time in secs

70

80

90

100

Lr(k)

0 −1 −2

Lu(k)

0 −0.01 −0.02

τk

8 6 4

Fig. 12. Variation in gains Lx ,Lr , and Lu with delay τ for tracking performance

92

6

S. Srinivasan et al.

Conclusion

In this investigation time varying delays in communication channels were modeled using MRAN, based on experiments conducted with Modbus over TCP/IP network. Information on network conditions such as loading, communication length, contention ratio, and number of connected nodes along with experimental delays was used to train the MRAN. Once trained using network conditions, MRAN generates delay estimates that were used to design adaptive controller for NCSs subjected to time-varying delays. MRAN-based estimation yields a learning based framework for estimating delays and, therefore, is more suitable for real-time control. The adaptive controller uses LQR approach to compute the gains and as a result, optimality is guaranteed by construction. Then, conditions for stability of the proposed controller were derived. The results were illustrated using two simple simulation examples along with delay estimates from MRAN. Results indicate that the adaptive controller modifies its gain based on time-varying delays in communication channels to meet the performance requirements. Adaptive controller by addressing packet-loss and implementing the adaptive control in real-time industrial pilot are the future prospects of this investigation. Acknowledgments. This work was supported by European Union through European Regional Development Fund and the Estonian Research Council grant PUT481.

References 1. Hespanha, J.P., Naghshtabrizi, P., Xu, Y.: A survey of recent results in networked control systems. Proc. IEEE 95(1), 138 (2007) 2. Seshadhri, S.: Estimation and design methodologies for networked control systems with communication constraints. Ph.D. thesis, NIT-Trichy (2003) 3. Zhang, L., Gao, H., Kaynak, O.: Network-induced constraints in networked control systems survey. IEEE Trans. Ind. Inform. 9(1), 403–416 (2013) 4. Seshadhri, S., Ayyagari, R.: Platooning over packet-dropping links. Int. J. Veh. Auton. Syst. 9(1), 46–62 (2011) 5. Qu, F.-L., Guan, Z.-H., Li, T., Yuan, F.-S.: Stabilisation of wireless networked control systems with packet loss. IET Control Theory Appl. 6(15), 2362–2366 (2012) 6. Srinivasan, S., Buonopane, F., Vain, J., Ramaswamy, S.: Model checking response times in networked automation systems using jitter bounds. Comput. Ind. 74, 186–200 (2015) 7. Balasubramaniyan, S., Srinivasan, S., Buonopane, F., Subathra, B., Vain, J., Ramaswamy, S.: Design and verification of cyber-physical systems using truetime, evolutionary optimization and uppaal. Microprocess. Microsyst. 42, 37–48 (2016) 8. Tzes, A., Nikolakopoulos, G.: LQR-output feedback gain scheduling of mobile networked controlled systems. In: Proceedings of the 2004 American Control Conference, 2004, vol. 5, pp. 4325–4329. IEEE (2004) 9. Godoy, E.P., Porto, A.J., Inamasu, R.Y.: Sampling time adaptive control methodology for can-based networked control systems. In: 2010 9th IEEE/IAS International Conference on Industry Applications (INDUSCON), pp. 1–6. IEEE (2010)

Adaptive Controller for Networked Control Systems Subjected

93

10. Piltz, S., Bjorkbom, M., Eriksson, L.M., Koivo, H.N.: Step adaptive controller for networked MIMO control systems. In: 2010 International Conference on Networking, Sensing and Control (ICNSC), pp. 464–469. IEEE (2010) 11. Bj¨ orkbom, M., et al.: Wireless control system simulation and network adaptive control. Ph.D. Dissertation, Dept. of automation science and technology, School of Science and Technology 12. Voit, H., Annaswamy, A.: Adaptive control of a networked control system with hierarchical scheduling. In: American Control Conference (ACC), 2011, pp. 4189– 4194. IEEE (2011) 13. Marti, P., Yepez, J., Velasco, M., Villa, R., Fuertes, J.: Managing quality-of-control in network-based control systems by controller and message scheduling co-design. IEEE Trans. Ind. Electron. 51(6), 1159–1167 (2004). doi:10.1109/TIE.2004.837873 14. Loden, N.B., Hung, J.Y.: An adaptive PID controller for network based control systems. In: 31st Annual Conference of IEEE Industrial Electronics Society, 2005, IECON 2005, pp. 6–pp. IEEE (2005) 15. Tahoun, A., Hua-Jing, F.: Adaptive stabilization of networked control systems. J. Appl. Sci. 7(22), 3547–3551 (2007) 16. Chunmao, L., Jian, X.: Adaptive delay estimation and control of networked control systems. In: International Symposium on Communications and Information Technologies, 2006, ISCIT 2006, pp. 707–710. IEEE (2006) 17. Xu, H., Jagannathan, S., Lewis, F.L.: Stochastic optimal control of unknown linear networked control system in the presence of random delays and packet losses. Automatica 48(6), 1017–1030 (2012) 18. Watkins, C.J.C.H.: Learning from delayed rewards. Ph.D. thesis, University of Cambridge (1989) 19. Luck, R., Ray, A.: Delay compensation in integrated communication and control systems: part II-implementation and verification. In: American Control Conference, 1990, pp. 2051–2055. IEEE (1990) ¨ uner, U.: ¨ Closed-loop control of systems over a communications 20. Chan, H., Ozg¨ network with queues. Int. J. Control 62(3), 493–510 (1995) 21. Nilsson, J., et al.: Real-time control systems with delays. Ph.D. thesis, Lund Institute of Technology Lund, Sweden (1998) 22. Seshadhri, S., Ayyagari, R.: Dynamic controller for network control systems with random communication delay. Int. J. Syst. Control Commun. 3(2), 178–193 (2011) 23. Ghanaim, A., Frey, G.: Modeling and control of closed-loop networked PLCsystems. In: American Control Conference (ACC), 2011, pp. 502–508. IEEE (2011) 24. Srinivasan, S., Vallabhan, M., Ramaswamy, S., Kotta, U.: Adaptive LQR controller for networked control systems subjected to random communication delays. In: American Control Conference (ACC), 2013, pp. 783–787. IEEE (2013) 25. Srinivasan, S., Vallabhan, M., Ramaswamy, S., Kotta, U.: Adaptive regulator for networked control systems: Matlab and true time implementation. In: 2013 25th Chinese Control and Decision Conference (CCDC), pp. 2551–2555. IEEE (2013) 26. Vallabhan, M., Seshadhri, S., Ashok, S., Ramaswmay, S., Ayyagari, R.: An analytical framework for analysis and design of networked control systems with random delays and packet losses. In: Proceedings of the 25th Canadian Conference on Electrical and Computer Engineering (CCECE) (2012) 27. Lian, F.-L., Moyne, W., Tilbury, D.: Optimal controller design and evaluation for a class of networked control systems with distributed constant delays. In: Proceedings of the 2002 American Control Conference, 2002, vol. 4, pp. 3009–3014. IEEE (2002)

94

S. Srinivasan et al.

28. Yingwei, L., Sundararajan, N., Saratchandran, P.: Identification of time-varying nonlinear systems using minimal radial basis function neural networks. In: IEE Proceedings of the Control Theory and Applications, vol. 144, pp. 202–208. IET (1997) 29. Platt, J.: A resource-allocating network for function interpolation. Neural Comput. 3(2), 213–225 (1991) 30. Kadirkamanathan, V., Niranjan, M.: A function estimation approach to sequential learning with neural networks. Neural Comput. 5(6), 954–975 (1993) 31. Yingwei, L., Sundararajan, N., Saratchandran, P.: A sequential learning scheme for function approximation using minimal radial basis function neural networks. Neural Comput. 9(2), 461–478 (1997) 32. Yingwei, L., Sundararajan, N., Saratchandran, P.: Performance evaluation of a sequential minimal radial basis function (rbf) neural network learning algorithm. IEEE Trans. Neural Netw. 9(2), 308–318 (1998) 33. Seshadhri, S.: Sampled-data model for ncss with delay less than sampling time (2012). http://www.mathworks.com/matlabcentral/fileexchange/37875/. Accessed 21 July 2013

Aspects Regarding Risk Assessment of Human Body Exposure in Electric and Magnetic Fields Marius Lolea and Simona Dzitac(&) Department of Energy Engineering, University of Oradea, Oradea, Romania [email protected], [email protected] Abstract. The paper gives an idea of attaching the risk of electromagnetic field exposure in the general population and employees with potential health hazards that are related to the different sensitivity of the human body. The difference between the two categories of people, is that in the first case, the intersection with the devices which generating electromagnetic field is optional both in duration, number or distance, and in the second case this option is canceled due to duties of services that give the obligation of employees to stationed in areas with such types of devices. Even if there are thresholds exposure limit values for electric and magnetic field quantities considered hazardous to health, referred to in legal settlements can not say with certainty that human sensitivity and is given the same thresholds for different age and sex categories. These fit rather with probability theory. Therefore the work is appreciated and that the risk to health or human behaviors generated by electromagnetic field exposure related to probability theory and treatment of this risk is therefore using a probabilistic model.

1 Introduction The multiple types of specific diseases of modern civilization and the large number of reported cases led to the idea of search for new cause increased risk of their generation. Among the causes were included and electromagnetic field exposure. Fears are amplified by the existence of numerous electromagnetic devices that help to increase life comfort and technical progress. Epidemiological and laboratory studies, conducted at worldwide lavel have highlighted several problems attributed field with different degrees of certainty. The results of these studies are controversial [5, 9, 11]. However, national rules limiting the exposure of persons, the maximum limit values are provided for parameters such as electric and magnetic field and the exposure so that the effects are not harmful to health [6, 7]. The interplay between electromagnetic fields and living matter are very different and depend on the frequency of electromagnetic waves, the amplitude of characteristic parameters of the field and proportion of energy transmitted to the body [5, 9–12]. Biological implications of exposure to magnetic fields against human body, is can study methodological with a cybernetic or an analytical method [5]. Cybernetic method consider the biological system as a black box characterized input quantities, output sizes and connections between them. Input values are given by electromagnetic field parameters. Output sizes are effects on the living organism. Connections between the sizes of input and output are formed from interaction mechanisms which are obtained on experimental basis. The analytical method is based © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_9

96

M. Lolea and S. Dzitac

on mathematical modeling of biological systems. The phenomena that occur at the molecular level to the application of an electromagnetic field are: electronic excitation, the polarization, the appearance of electrical forces, thermal effects, dipolar interactions, etc. Exist, circuit diagrams of the living cell developed and mathematical models to study the electromagnetic interaction at anatomic level [5]. In a few situations electromagnetic field impact with human body it is a risk factor occupational diseases [4]. The purpose of the paper is to assess the level of human exposure through measurements values for the electromagnetic field generated by electrical power networks equipments and risk of exceeding normal levels considered dangerous of these sizes. For this second case apply a probabilistic model [1–3].

2 Probabilistic Methods from Risks Assessment The risk associated with occurrence of a certain event comes down, in the simplest form, as the product of the probability of that event and the consequences of that event. For example, for a given level of risk desired, as the likelihood of an event decreases the potential consequences products may increase. All the events (states), through which a physical system during the experience is called field events. By associating a field event of an experiment may be obtained depth information on knowledge of the phenomenon studied [1, 3]. The probability transferred in quantitative and intuitive plan of numbers, a concrete situation pattern of events, allows some comparison of events. In this way, the probability is a measure of risk. The concepts of probability and probability finite field can be presented as axiomatic form. By Kolmogorov, probability is defined by the following axioms: • each event A 2 K, of the field events K, is associated with a positive number P {A}, called the probability of the event; • P {E} = 1 is the probability of the event is safe and has a maximum value; • if two events A, B 2 K, are incompatible, then: {P} = P A [ B {A} + P {B}. The concepts of probability of appearance p, and non-appearance probability r of an event for which p + r = 1, allow to judge between safety and risk that there is a complementary relationship. On the basis of the axioms of Kolmogorov, it follows that R(p) + S(r) = 1, where R(p) is the probability of risk corresponding to an event, and S(r) - additional safety risk. The relation: R(p) + S(r) = 1 reveal that an increased risk involving a decrease of safety and a risk reduction involve an increases of the safety, the two sizes having values in the range [0, 1]. In the general case it can write: 0 < R(p) < 1 sau 0 < S(r) < 1, after which, in particular: If R(p) ! o, it can write S(r) ! 1 and inverse. If danger determines the risk, then the risk can be defined as a measure of danger P (p) and its extent permitted limit value: R(p) = limn!1 PðnÞ ¼ P(p)

Aspects Regarding Risk Assessment of Human Body Exposure

97

For the particular case of the action of the electromagnetic field on the human body either by direct coupling or by electromagnetic radiation, the danger is delimited by limit values standardized of field quantities considered harmful. Thus, assessing the probability of exceeding the normal levels that are considered generating negative health effects, is a step in assessing risk from exposure. The risk here can be defined as the product of probability of exceeding the exposure limit values for E and B and that the consequences of these exceedances, the latter being far damage human life or health. Following graphs below of normal distributions, we can estimate the risk from producing an unwanted event by leaving the chosen interval of significance [8] (Fig. 1).

Level of significance 1-α

Level of significance 1-α

Risk α

Risk α/2

Risk α/2

Interval of significance

Interval of significance

θα/2

θα

θα/2

Fig. 1. The link between the meaning range and of probability: a - to a one-sided risk; b - for a bilateral risk (adapted from [8])

To a certain repartition function, meaning the connection between interval and level of significance, it is given by relation: Pðh\ha Þ ¼ 1  a;

ð1Þ

for a unilateral risk, or by relation: P(h1a2 \h\ha2 Þ;

ð2Þ

For a bilateral risk, where: 1 − a is the level of significance; ha − hb = interval of significance; a – the risk. Assessed risk the consequences of exposure to electromagnetic field is generated type as unilateral effects considered dangerous to humans is only upside quantities of electric and magnetic field above their normal intake. A typical case of bilateral risk is generated by electromagnetic disturbances such as voltage fluctuations. Whether the voltage drops below a certain threshold, as in dips or rises above the maximum value that this service became Surge, risks in product supply electricity to consumers are double effect.

98

M. Lolea and S. Dzitac

3 Risk Assessment Generated to Humans Exposured in Electromagnetic Field Generated by Power Grids Assessing the risk of exceeding the permissible values for magnetic induction (B) and electric field (E) is performed by comparing the density function of considered size distribution (f(B) and/or (E)) with the permissible limit value of these sizes set of regulations in the field of workers’ exposure to electromagnetic fields and/or the general population [2, 3]. For calculations follow the next steps: (a) shall be taken from measurements the values of magnetic induction B and electric field strength E at various locations; (b) is calculated for each location, the arithmetic mean and standard deviation of the values taken at the measuring points; (c) Considering that values have a normal distribution, determine the probability density function expression (f(B)/f(E)). The form of this function is [3]: 1 Mm 2 1 f(M) = pffiffiffiffiffiffi  e  2 ð r Þ 2p  r

ð3Þ

where: m – the mean of measured values; r – standard deviation of measured values; M = field sizes considered: {E, B}. Measured values of the magnetic induction and electric field strength is considered that follow a normal distribution. Calculation was developed in Mathcad application, version 14.0. The calculation algorithm initiated by the authors has the following form: • Defines the Matrix of the maximum measurements of magnetic induction B and electric field strength, E. It has the following tabular form for B, in the first case: Measuring data

B1 1

2

1

11.4

2

12.6

3

13.5

4

...

• Number of datasets from matrix B1: ! NCol := 1 The limit value for B: ! Blim := 100

3

Aspects Regarding Risk Assessment of Human Body Exposure

99

The computing vector, has the form:

The elements of the vector above are: DD1 = mean, DD2 = standard deviation, DD3 = the risk; No. of column inside matrix B1, k: ! k := 1   f(B): = dnorm B,ðDD1 Þk ; ðDD2 Þk ! Density distribution function, for B;   f1(B): = dnorm B; ðDD1 Þk ; ðDD2 Þk ðB [ BLim) ! Marks the risk;

• Drawing graph to distribution function, finally. The algorithm applies to workers of Hydropower plants(HPP) and for the general population. Through the activities undertaken and functional characteristics of the electricity networks they serve workers are exposed to electromagnetic fields generated by them. The electromagnetic environment is formed due to the operation of all sources of a given site or a particular territory. Therefore for each location we need to consider

100

M. Lolea and S. Dzitac

all sources of electromagnetic field. Measuring devices used were: CA 42 Field meter and SPECTRAN 5035, both equipped with internal probes of magnetic field and external for electric field. Displaying values is in V/m for E and in lT for B. Measuring step is noted by dm and is chosen as equal to 2 m, to detect with high accuracy, the maximum values of quantities of electric and magnetic field. With measured values will be calculated the averages, standard deviations and risk probability [1–3]. For the general population analyzes a portion of a high voltage overhead power lines (PL), which crosses the city of Oradea through a crowded area. In the case of power lines(PL) measurements have been performed under conductors, in line axis. Technical characteristics of the PL for the aperture considered, are: circuit type d.c., Un = 110 kV, connection between substations Oradea Sud – Oradea Vest, the portion of Auchan supermarket, on Radu Enescu street, type poles: STC, aperture lenght = 126 m, composite insulators, conductors type: OL-Al 185/32 mm2, gauges: cond.1 = 13.7 m, cond.2 = 13.4 m, cond.3 = 14.05 m, cond.4 = 13.17 m, cond.5 = 14.25 m, cond.6 = 13.76 m; number of measurements: 126/2 = 63 at measure step considered. On the HPP Tileagd case, is exemplified with the values of magnetic induction B, taken around of the generators excitation, where operational staff have access, expressed in lT. In the case of generators the measurements were carried out on a circular path around them. Threshold values within normal limits, are [6, 7]: Elim = 5000 V/m for the general population, Elim = 10000 V/m for employees; Blim = 100 lT for the general population; Blim = 500 lT for employees. Allowable limit values were periodically adjusted by their decreasing, due to the fact that research into exposure domain have been perfected over time. Therefore, to consider the Gaussian variations of their, is appropriate. With considered fixed limit values, Fig. 2 shows the risk graph for the aperture analyzed and in Fig. 3 shows the risk graph for operators from generators hall of HPP Tileagd, belonging to Bihor Power System. With blue color was marked risk area. 3×10

2.25×10

f (E) f1(E)

1.5×10

7.5×10

−4

8×10

−4

6×10 f (B)

−4

f1(B)

−5

4×10

2×10

0 0

2×10

3

4×10 E

3

6×10

3

8×10

3

−3

−3

−3

−3

0

0

87.5

175

262.5

350

B

Fig. 2. The graph of risk: a - with values of E for PL 1/110 kV, aperture 1; b - with values of B for considered case inside HPP Tileagd

Aspects Regarding Risk Assessment of Human Body Exposure

101

Fig. 3. Setting data for considered case in MathCad software

For evidence of common risk corresponding of the area under the intersection of the two curves are written equations corresponding to that point unknown p. For identification of solution is necessary to equating the two equations. How the two points are at the same level on the chart at the intersection of the two curves density distribution, they must have the same coordinates so p1(x1, y1) = p2(x1, y1) and can write:   1 p  m1 2 1 for f(p1 Þ ¼ pffiffiffi  e 2 1 r1 2pr1

ð4Þ

 2 1 12 p2  m2 for f(p2 Þ ¼ pffiffiffi e radm 2pradm

ð5Þ

For the case in which the standard deviations for the two sets of data, ranging to Gaussian distribution are equal, result: r1 ¼ radm ¼ r;

ð6Þ

and because the two points coincide on chart, namely p1 = p2 = p, by equalizing the two expressions follows:     1 1 1 p  m1 2 1 p  m2 2 pffiffiffi  e2 ¼ pffiffiffi  e2 r r 2pr 2pr

ð7Þ

102

M. Lolea and S. Dzitac

By logarithm, will get: 

1 p  m1 2 1 p  m2 2 ¼ 2 2 r r

ð8Þ

And after simplifying and solving the resulting equation, the unknown being selected value will be calculated with the following expression: p=

m1 þ m2 ; 2

ð9Þ

then, in case of the means of the written program running in MathCad will get: p = 70.206, with m1 = 40.412 and m2 = 100. For the other calculation hypotheses, in which radm ¼ 12 r and radm ¼ 14 r resulting the equalities: 1 1 pm1 2 1 1 For case 2 : pffiffiffi  e2ð r Þ ¼ pffiffiffi r  e2 2pr 2p  2 1 1 pm1 2 1 1 For case 3 : pffiffiffi  e2ð r Þ ¼ pffiffiffi r  e2 2pr 2p  4

pm 2 2 r=2

pm 2 2 r=4

ð10Þ

ð11Þ

Which have the next solutions: For Eq. (10):

p1;2 ¼

4m2  m1  2

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 4ðm1  m2 Þ2 þ 6rln2 3

ð12Þ

and in case of Eq. (11):

p1;2 ¼

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 16m2  m1  2 16ðm1  m2 Þ2 þ 30r2 ln4 15

ð13Þ

To generate common area beneath the two curves, the left and right of the point of intersection called risk area, perform the necessary settings in Mathcad. These are indicated in Fig. 3. In Fig. 4 is shown in green color, the resulting risk in case the two standard deviations are equal. In case of the measured magnetic flux density values exemplified for synchronous generators hall of o HPP Tileagd, they were chosen radial directions for further circular and concentric routes around excitation systems. Since the operators do not have to stand throughout the service duration near excitation systems, the acceptable value exposure to magnetic induction is considered the value of Blim = 100 lT. In Figs. 5 and 6, r, r2 and r3, are the standard deviations for another two considered cases.

Aspects Regarding Risk Assessment of Human Body Exposure

103

p 70.206 Fig. 4. Risk graph in the second case considered

case 2 σ,σ2

m1

m2 Risk zone

Fig. 5. The distribution of magnetic induction admissible with standard deviation in portion of 0.25% from r of a real values measured

case 3 σ,σ3

m3

Risk zone

Fig. 6. The distribution of magnetic induction admissible with standard deviation in portion of 0.50% from r of a real values measured

104

M. Lolea and S. Dzitac

Algorithms presented in the paper, apply for other power plants in Bihor County. Algorithms presented in the paper, apply for other power plants in Bihor County. For the first scenario considered when considered dangerous occupational exposure limit values are considered fixed, calculated parameters are summarized in Tables 1 and 2. Probability of exceeding the admissible limits, allows ranking of these facilities in terms of the risk from exposure to the population and employees in the energy sector. This is shown in Fig. 7.

Table 1. Calculated parameters values for B, in case of HPP Crt. No. 1 2 3 4 5 6

Objective of the measurements HG hall -HPP Fughiu HG hall HPPSăcădat HG hall - HPP Tileagd HG hall - HPP Lugașu HG hall - HPP Munteni HG hallo – HPP Remeți

Measurements number 246

Mean (M) 22.5

Standard deviation (r) 0.094554

Overtaking probability (risk/Pr) 0.047421

256

37.5

0.128673

0.023425

560

25.7

0.137553

0.039424

580

35.8

0.154553

0.023423

380

22.1

0.188553

0.015976

380

28.3

0.174553

0.024436

Table 2. Calculated parameters values for E, in case of PL Crt. No. 1 2 3 4 5 6

Objective of the measurements PL 1 PL 2 PL 3 PL 4 PL 5 PL 6

Measurements number 1246 1756 2560 2180 980 1380

Mean (M) 22.5 37.5 25.7 35.8 22.1 28.3

Standard deviation (r) 0.134554 0.124673 0.134553 0.184553 0.184553 0.154553

Overtaking probability (risk/Pr) 0.037421 0.057423 0.038424 0.033423 0.046475 0.033436

Figure 7 shows the ranking of categories of people according to risk exposure caused by the electric and magnetic field in the two objectives analyzed.

Aspects Regarding Risk Assessment of Human Body Exposure

105

0.06 HPP 6

PL 3

0.04

HPP 3

PL 5

0.02

HPP 1

PL 2

0.00 PL ranking

HPP ranking

Fig. 7. The Graph of risk for exposure in electric field. PL and HPP ranking

4 Conclusions In the context of scientific uncertainty regarding the effects of exposure to electromagnetic fields advisable to adopt the precautionary approach by implementing administrative measures, inform and train the population, especially those at risk of exposure sources and not least support from decision makers, on the development of an adequate logistics structure from monitoring electromagnetic fields and their effects in the long term. The calculation model which is adopted, not give information about risk consequences but, based on the distribution of the measured values enable the assessment of the likelihood of exceeding the normal threshold for electric field intensity electric and magnetic field induction on exposure to electromagnetic field of workers from the power system or the general population. Exposure Risk assessment is done by comparing the results of calculation of probability values exceeded the normal values of field sizes for all staff categories analyzed. Standard deviations and default the dispersions of measured values, are more pronounced where the measuring points were distributed over large areas with a great variety of equipment and devices that work with large differences between voltages and currents service. However, if one considers that there are sensitivities and different reactions bodies in terms of exposure to electromagnetic field as suggested in the literature, along with the assumption accepted the job as the limit values allowable quantities Field pursue distribution normal, then the risk of negative effects on the human body increases. This is evident because the overlap of the two distributions is more pronounced in the lower levels of the two categories of values, measured and admissible. Then probabilistic model adopted is justified especially if it is found that the values of magnetic induction and electric field strength, results of measurements, follow a normal distribution.

106

M. Lolea and S. Dzitac

References 1. Felea, I., Coroiu, N., Stoica, I., Dubău, C.: Evaluarea unor elemente de risc pentru personalul de exploatare al rețelelor electrice. Analele Universității din Oradea, fascicula de energetic, vol. I, pp. 37–50. Secțiunea II/Electroenergetică (2001) 2. Felea, I., Secui, C., Muşet, A.: Evaluation of risk affecting the exposed human organism in operational electromagnetic field. In: The 6th International workshop of Electromagnetic Compatibility CEM, Constanţa, Romania, 12–14 November 2009 3. Felea, I.,Coroiu, N., Secui, C.: Asupra riscului de expunere în câmp electromagnetic. Simpozionul Național Siguranța în funcționare a Sistemului Energetic SIG, SC Electrica Deva, 26–28 September, vol II, pp. 402–410 (2001) 4. Virginia, M.: Câmpuri electromagnetice de joasă frecvenţă factori de risc profesional, Rev. Acta Medica Transilvania, II(1), 27–30 (2011) 5. Mihaela, M.: Bioelectromagnetism, Editura Matrix Rom, București (1999) 6. Guvernul României, Hotarârea Nr. 1136 din 30.08.2006 privind cerintele minime de securitate si sanatate referitoare la expunerea lucratorilor la riscuri generate de câmpuri electromagnetice 7. Guvernul României, Hotarârea Nr. 1193 din 29.09.2006 privind limitarea expunerii populatiei generale la câmpuri electromagnetice de la 0 la 300 GHz 8. Munteanu, T., Gurguiatu, G., Bălănuţă, C.: Fiabilitate şi Calitate în Inginerie electrică. Aplicaţii (ed.) Galati University Press (2009). ISBN 978-606-8008-25-7 9. Milham, S.: Dirty Electricity, Electrification and the Diseases of Civilization. Universe Star, New York (2012). ISBN 13:9781938908187, ISBN 10:193890818X 10. Aliyu, O., Maina, I., Ali, H.: Analysis of electromagnetic field pollution due to high voltage transmission lines. J. Energy Technol. Policy 2(7), 1–10 (2012). ISSN 2224-3232 (paper), ISSN 2225-0573 (online) 11. Otto, M., Muhlendahl, K.E.: Electromagnetic fields (EMF): do they play a role in children’s environmental health (CEH)? Int. J. Hyg. Environ. Health 210, 635–644 (2007). www. elseiver.de/ijheh 12. Hardell, L., Sage, C.: Biological effects from electromagnetic field exposure and public exposure standards. Biomed. Pharmacother. Rev. (2), 104–109 (2008). www.elseiver.com/ locate/biopha

A Communication Viewpoints of Distributed Energy Resource Ravish Kumar1 , Seshadhri Srinivasan2(B) , G. Indumathi3 , and Simona Dzitac4 1

2

ABB Corporate Research, Bangalore, India [email protected] International Research Center, Kalasalingam University, Virudhunagar, India [email protected] 3 Cambridge Institute of Technology, Bangalore, India [email protected] 4 Universtatea Din Oradea, Oradea, Romania [email protected]

Abstract. Smart Grid provides a flexible and powerful framework to integrate Distributed Energy Resources (DER). DER consisting of energy sources (such as renewable, conventional, storage) and loads. Based on application, there can be different modes of operating. It can operate in parallel with, or independently from, the main grid. Because of social, economic, and environmental benefits, the demands of DER system is increasing gradually. Combination of DERs and Loads with control units forms a miniature grid which is known as Microgrid. Integrating of DERs to the Microgrid system has two aspects: electrical and communication. Electrical integration has their own challenges which are addressed in IEEE 1547 standard. However, challenges related to communication and information integration of multi vendor DERs system is still open. Currently, IEC 61850-7-420 is the only standard which defines information model for DERs, but it is not completely accepted by DER manufacturers. On the other side, IEC 61850 mapped communication protocols like MMS, GOSSE, DNP3 etc. are substation protocols and may not suitable to fulfill DERs to DERs communication demands. In this paper, we provide an overview of Smart Grid conceptual framework and highlight DER scope. We discuss DER system components, various applications and required communication features. We review IEC 61850 mapped communication protocol MMS and identify the limitation for using in DERs communication.

1

Introduction

The traditional electrical grid is an interconnected network for delivering electricity from suppliers to consumers. It consists of generating stations that produces electrical power, high-voltage transmission lines that carry power from distant sources to demand centers, and distribution lines that connect individual customers. The growing demand of electricity led to increase numbers of electrical c Springer International Publishing AG 2018  V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8 10

108

R. Kumar et al.

grids deployment located at different geographical location. The electrical grid is a broadcast grid networks, where a few central power generator provides all the electricity production to the industrial and residential customer via a large network of cables and transformers. Electrical grid structure can summarized as follows: one-way energy flow from central station generators, over a transmission network, through substations onto distribution systems, over radial distribution circuits to end-use customers. To address the increasing energy demands, the traditional electrical grids being enhanced by adding more bulk generators to produce more energy. Bulk generators mostly use fossils fuels for power generation. On this earth, there are limited fossils fuels. Rapid consumption of limited natural resources must be avoided for future sustainable energy generation. To meet the future demands of electricity it is being focused to shift traditional grid generating power from fossils fuels to renewable resources [1]. The renewable resources are solar, wind, small hydro, tidal, biogas etc. The future of electric grid is Smart Grid (SG) [1]. The objective of SG is to improve energy efficiency and reduce carbon footprint at all the stages of energy flow: Bulk Generation, Transmission, Distribution and Consumption. SG vision is to provide an open framework and protocol for information and power exchange seamlessly. Unlike traditional electric grid, which is centralized grid, the SG is decentralized grid. It allows to integrate power sources at distribution level. The end consumer can be Prosumers, which acts as electricity consumer and producers at the same time. Prosumer have their own small Distributed Energy Resources (DERs) for electricity generation. Based on local usage, the additional generated power can be supply to utility grid. Presently, most of the prosumers have limitation, they can not supply surplus power to utility grid because of insufficient electrical and communication infrastructure [2]. The surplus overproduced electricity is just thrown away, if prosumer’s energy storage are full. Lack of infrastructure support causes waste of energy. The benefits of the DER integration in smart grids has been discussed in [12–15]. Because of social, economical and environment benefits, DER technology is evolving, new DER systems are getting deployed across the globe. However, research are still lacking for providing standard communication interface for DERs [3] system. Real time information exchange is very crucial to ensure reliable power supply. Presently, most of the DER system communicates with other system either on proprietary protocols or basic protocols like Modbus. There is no standard communication protocol and interface are recommended for DER system by International Standard organization. DER system is considered as an alternative of main power grid. New applications are developed to connect with DER system to enable new features. Lack of standard interfaces and protocols limitation to provide interoperable among different manufacture devices and system. International standard IEC 61850 [4] is known and worldwide accepted standard for Substation Automation. It decouples the application data model from communication protocols and provide flexibility for extending this standard for any application. Because of this flexibility, addition of new domain is always

A Communication Viewpoints of Distributed Energy Resource

109

possible with IEC 61850. It is also being consider as a prime mover of SG framework. Since, it brings flexible and open framework for monitor, control, and measurement applications. The IEC 61850 7-420 [5] standard defines information model for DER, that describe what information and semantics to exchange with other DER system. DER communication with existing IEC 61850 mapped protocols such as MMS, DNP3, GOOSE can be implemented. However, from DER functionality point of view, many of the communication needs may not be fulfilled by these protocols. For example, auto device registration and plug and play, session less communication among DERs, etc. Authors in [11] investigated the role of OPC UA and IEC 61850 integration for enabling smart grids and demonstrated how OPC UA can provide communication medium for smart grid elements. Interoperability is the key aspects DER system. One DER system must be interoperable with other DER system. Interoperability is an ability of two systems to communicate and share information with each other. The communication aspect of devices can be divided into four aspects [5]: Information modeling (the type of data to be exchanged-nouns), Service modeling (the read, write, or other actions to take on data-verbs), Communication protocols (mapping the noun and verb models to actual bits and bytes), and telecommunication media (fiber optics, radio system, wire, and other equipment). Currently, IEC 618507-420 is the only standard which defines information model and data model for DER system. However, the existing IEC 61850 mapped protocol seems not to be suitable for DERs communication. Most of the DER system vendors does not support IEC 61850 communication protocols in the DER system. They use their proprietary for simplicity and ease of development. Manufacturers of DER system would like to continue DER communication with their proprietary protocols. As we mentioned earlier, IEC 61850 standard is designed for Substation Automation, where system configurations is static and preconfigured. Therefore, existing IEC 61850 protocols may not be suitable for DER system communication. Yoo et al. [6] has shown IEC 61850 communication architecture based Microgrid communication and only considered static DER system configuration. Suˇci´c et al. [7] highlighted the IEC 61850 communication protocol limitations for DER system and suggested Service Oriented Architecture for DER system. In this paper, we discuss an overview of Smart Grid and DER. Further, we discuss different applications of DER system and provide the required communication viewpoints. The rest of the paper is organize as follows. Section 2 provide an brief overview of Smart Grid Architectural Model Framework and discuss different domains and zones. In Sect. 3, we discuss the Distributed Energy Resources system, its application, and communication architecture. Section 4, we have evaluated IEC 61850 communication protocols for DER communication. Section 4 discuss challenges for implementing IEC 61850 communication protocols in DER system. And, in Sect. 5 conclude the whole paper.

110

2

R. Kumar et al.

Smart Grid Conceptual Framework

Smart Grid (SG) is a combination of different technologies and can be seen as complex system. The key enabler of SG framework is an open communication and integration framework, which allows bidirectional information and electricity flow. Figure 1 shows a conceptual Smart Grid Architecture Model(SGAM) proposed by National Institute of Standard and Technology [1].

Fig. 1. SGAM conceptual framework model [1]

SGAM framework consists of Business layer, Function layer, Information layer, Communication layer, and Component layer. An interoperability aspect must be satisfied among the layer for successful Smart Grid operation. The Smart Grid system is divided into two aspects Electrical process and Information management. In the SGAM framework, domains represent electrical process and Zones represent the information management. Domains physically related to the electrical grid components: Bulk Generation, Transmission, Distribution, Distributed Energy Resources (DERs), and Customer premises. It is arranged according to the electrical energy conversion flow. Zones are related to power system management. Zone reflects the hierarchical level based on aggregation and functions separation of each hierarchical level. DER is considered as basic building block for Smart Grid system. Since, it provides ancillary distributed energy resources at distribution level, close of end consumers. The importance of DER are being recognized due to its social, economic, and environmental benefits. Growing number of DER demand is estimated to reach USD 34.94 Billion by 2022, at a CAGR of 10.9% between 2016 and 2022 [8].

A Communication Viewpoints of Distributed Energy Resource

3

111

An Overview of Distributed Energy Resources System

DER plan is consisting of different energy sources (e.g. renewable, conventional, storage) and loads. It is capable of operating in parallel with, or independently from, the main grid. DER system is more than back-up power source. Figure 2 shows an example of DER system.

Fig. 2. Distributed Energy Resources system example

DER plant operates in two modes: Grid Connected (grid mode) and Grid Isolated (islanding mode). In the grid mode, DER plan is connected to utility grid Common Coupling Point (CCP). It is being controlled based on utility grid power supply and cost factor behavior. For example, if DER is generating energy using renewable sources, then the load will use energy from and optimize the utility grid power in order to save overall cost. While in the case of islanding mode, DER is not connected to utility grid. It acts to remote power sources. Load is completely relied on DER. Based on the condition the transition takes place between, grid to islanding mode, and islanding to grid mode. The fundamental operational requirements of DER plant to provide stability of frequency and voltage for ensuring optimal power flow to the load. Grid to islanding mode requires transition and stabilization for minimal load shedding and distribution. Islanding to grid mode requires resynchronization and minimum impact for sensitive loads during transient periods. Each of these operations are highly dependent on underlying communication infrastructure. The growing demand of DER system increased the complexity of multivendor DER elements interconnection, since most of the DER plant vendors use their own proprietary communication protocols.

112

R. Kumar et al.

Typically, DER plant may have three level of controller. Primary, secondary and tertiary controller [9]. Primary controller is machine’s (such as diesel generator) local controller. The secondary controller monitors and controls the primary controller based on the real time to operate and control generating machine behavior. Tertiary controller is used to controller power flow among DER systems located at different location. In the below subsection DER applications example are discussed to explain DER operating modes. 3.1

DER Applications

DER provides an alternative local power source in parallel or independently from main grid. DER system can be simple or complex as per the application needs and use case. The following are the application scenarios of DER application. 3.1.1 Back up for Poor Power Quality of Main Grid Poor power quality refers inability of main grid to provide stable power supply with any fluctuations. Power quality of grid strongly depends on the load characteristic and transmission and distribution grid infrastructure. Long transmission line with asymmetric loads can easily influence power quality, which can results into unbalanced voltage, harmonic, in power supply. Poor power quality may lead to power outage, if demands lead to power generation capacity. In such cases, DER system at the consumer site plays an important role. Depending on the field of application, military, industrial, commercial, or residential, power quality requirement of different DER system can be deployed to provide back up when there is poor power quality supply from main grid. 3.1.2 Natural Disasters Natural disasters such as earthquake, tornados, hurricanes, and tsunamis may damage power grid transmission and distribution network and the area which are not directly affected by natural disasters may suffers from power outage for a week to months if damage is severe. Since DER system is not dependent on power supply of the power grid, the immediate construction of DER system can provide an alternate power supply. 3.1.3 Power Grid Failure Large power grid, chances of disturbances due to cyber-attack, terrorist attack, human error, natural incidents, etc. are very high. The cascading effect of one unit failure can end up power outage in larger region. A grid isolated DER systems can be an alternative of Power grid failure in order to maintain power supply for crucial domains. 3.1.4 Addressing of Growing Power Demand The growing power demand results in unstable power supply if generation capacity lags the power demand. Upgrading of power grid generation is not an easy

A Communication Viewpoints of Distributed Energy Resource

113

process it requires huge investment and effort. Alternately, DER system is good candidate for satisfying growing demand of energy by deploying DERs near the high power demand zone. 3.1.5 Continuous Power Supply DER is the primary source of electricity and provides power supply to the load round the clock. 3.1.6 Uninterrupted Power Supply For this application, the crucial task of DER is very short start-up and ramp time to meet load demands. DER system should be stable and efficient. Typically, the operating hours of such DER system is less than two hours. 3.2

Der Communication Architecture

From the communication point of view, DER system can be divided into three layers. Layer 1 - Machine/Primary controller to Secondary controller communication, Layer 2, Secondary controller to Central supervisory control, and Layer 3, central supervisory controller to other DER systems. These three layers are depicted in Fig. 3. Every generation unit such as PV array, wind turbine, diesel generator has its own primary controller unit, which locally controls the generator unit. Primary controller can start, stop and regulate the power generation. Primary controller interacts with secondary controller mostly using serial to analog interface. The job of secondary control is very crucial; it monitors the status of generator units and controls the physical aspects of generator based on vital statistics such voltage, frequency, circuitry health, etc. and broadcast the status information horizontally to other Secondary controller and vertically to Central Supervisory

Fig. 3. Distributed Energy Resources communication architecture

114

R. Kumar et al.

Control system. Communication between primary and secondary controller has highest frequency of message. High frequency of message of message is necessary for capturing and address transient behavior of the generators such as wind turbine, diesel generator etc. The message update cycle between primary controller and secondary controller is 10 ms, Secondary controller to secondary controller is 50 ms, Secondary controller to central supervisory controller is 100 ms, and, DER system to another DER system is more than 1 s. DER system will have advance features for ensuring autonomous operations. Some of these advance features are given below. 3.2.1 Automatic Discovery of DER System This feature allows DER system to discover new DER system in the network and synchronize their activities automatically. For example, if new wind turbine is connected to network, the existing DER system should be able to discover without having any prior inform of DER. 3.2.2 Dynamic System Configuration This feature allows DER system element to reconfigure themselves automatically based on the overall plant status in order to provide reliable, efficient and good quality of power supply. 3.2.3 Plug and Play This feature allow new DER element to be part of network automatically. DER system will be configured automatically by Central Supervisory Control system. 3.2.4 Semantic Data Interpretation It is expected that dataset of DER system should be semantically interpreted by other DER system. This can be achieved if all the DER system support IEC 61850-7-420 information model.

4

Analysis of IEC 61850 Communication Protocols for DER System

The IEC 61850 standard is specially designed for Substation Automation domain. The communication model is designed in way the application specific data exchange is independent from communication model. Therefore, IEC 61850 can be extended to other domains as well. IEC61850 standard allows the integration of all protection, control, measurement and monitoring functions by one common protocol. It provides the means of high-speed substation applications, station wide interlocking and other functions which needs intercommunication between IEDs. The advantage of using IEC 61850 standard is that it ensures the interoperability and information interchangeability among multivendor devices. Device from any vendor compliant to IEC 61850 standard can communicates

A Communication Viewpoints of Distributed Energy Resource

115

Fig. 4. IEC 61850 based message exchange Table 1. Analysis of IEC 16850 mapped MMS protocol for DER DER required features

IEC 61850 support

Data modeling

Yes

Interoperability

Yes

Scalability

Yes

10–100 ms message transfer

Yes

Secure communication

No

Easy configuration

No

Dynamic configuration

No

DER auto discovery

No

Plug and play integration

No

Stable for low resource device

No

Point to point communication

Yes

Point to multi-point communication Yes Plug and play support

No

Less engineering effort

No

Time synchronization

Yes

TCP/IP

Yes

Wireless communication

No

Web service support

No

Protocol stack extensibility

No

File transfer

Yes

Publisher/subscriber communication No

with each other seamlessly. IEC 61850 standard defines an Abstract Communication Service Interface (ACSI) for information exchange and control among devices. The ASCI defined services are used for polling of data model information, reporting and logging of events, control of switches and functions. Peer to peer communication for fast data exchange between the feeder level devices

116

R. Kumar et al.

(protection devices and bay controller) is supported with GOOSE protocol. Voltage and current readings which are time synchronized are supported with Sampled Value communication model. Figure 4 shows the IEC 61850 based communication between two applications. Disturbance recordings are sent using File Transfer model. A common formal description code, which allows a standardized representation of a systems data model and its links to communication services is presented by the standard and its called as SCL (Substation Configuration Language) [10]. It covers all communication aspects according to IEC 61850. Based on XML schema, this code is an ideal electronic interchange format for configuration data, an equivalent of EDDL in Process Automation applications. The data model and the communication services (defined through ACSI) are decoupled from specific communication technology. This technology independence guarantees long term stability for the standard and opens up possibility to switch over to future communication technologies. Table 1 shows our analysis of IEC 61850 mapped MMS communication protocol for DER communication. We found MMS protocol may not be suitable for DER communication.

5

Conclusion

Distributed Energy Resources are an important Smart Grid building block. Combining multiple DERs unit forms a Microgrid which can be considered as miniature Smart Grid. Due to social, economic, and environment benefits, the demand of DERs deployment is increasing gradually. Currently DER system having interoperability issue. DER system of one manufacturer may not able to operate together because of their proprietary protocol use. In this paper, we have provided details of DER systems and it communication architecture. And provided communication view of point of DER system, which should be supported by standard protocols interface in coming future.

References 1. NIST Special Publication 1108R2. NIST Framework and Roadmap for Smart Grid Interoperability Standards, Release 2.0 2. Final report, Standards for Smart Grids. CEN/CENELEC/ETSI Joint Working Group 3. Cleveland, F.M.: IEC 61850-7-420 communications standard for distributed energy resource (DER). In: Proceedings of IEEE PES General Meeting, pp. 1–4 (2008) 4. International Standard IEC 61850: Communication networks and systems for power utility automation 5. IEC 61850-7-420 Ed. 1.0. Communication networks and system in power utility automation - Part 7-420: Basic communication structure - distributed energy resources logical nodes. March 2009. http://www.iec.ch 6. Yoo, B.-K., Yang, S.-H., Yang, H.-S., Kim, W.-Y., Jeong, Y.-S., Han, B.-M., Jang, K.-S.: Communication architecture of the IEC 61850-based micro grid system. J. Electr. Eng. Technol. 5, 605–612 (2011)

A Communication Viewpoints of Distributed Energy Resource

117

7. Suˇci´c, S., Havelka, J.G., Dragiˆcevi´c, T.: A device-level service-oriented middleware platform for self-manageable DC microgrid applications utilizing semantic-enabled distributed energy resources. Int. J. Electr. Power Energy Syst. 54, 576–588 (2014) 8. http://www.marketsandmarkets.com/Market-Reports/micro-grid-electronics-mar ket-917.html. Accessed Aug 2016 9. Kounev, V., Tipper, D., Grainger, B.M., Reed, G.: Analysis of an offshore medium voltage DC microgrid environment– part II: communication network architecture. In: IEEE PES T&D Conference and Exposition, pp. 1–5, Chicago, IL, USA (2014) 10. IEC 61850-6: Communication networks and systems for power utility automation – Part 6: Configuration description language for communication in electrical substations related to IEDs 11. Srinivasan, S., Kumar, R., Vain, J.: Integration of IEC 61850 and OPC UA for smart grid automation. In: 2013 IEEE Innovative Smart Grid Technologies-Asia (ISGT Asia). IEEE (2013) 12. Srinivasan, S., Kotta, U., Ramaswamy, S.: A layered architecture for control functionality implementation in smart grids. In: 2013 10th IEEE International Conference on Networking, Sensing and Control (ICNSC). IEEE (2013) 13. Verrilli, F., Gambino, G., Srinivasan, S., Palmieri, G., Del Vecchio, C., Glielmo, L.: Demand side management for heating controls in microgrids. IFAC-PapersOnLine 49(1), 611–616 (2016) 14. Gambino, G., Verrilli, F., Canelli, M., Russo, A., Himanka, M., Sasso, M., Srinivasan, S., Del vecchio, C., Glielmo, L.: Optimal operation of a district heating power plant with thermal energy storage. In: 2016 American Control Conference (ACC), pp. 2334–2339, July 2016. IEEE (2016) 15. Maffei, A., Srinivasan, S., Iannelli, L., Glielmo, L.: A receding horizon approach for the power flow management with renewable energy and energ storage systems. In: 2015 AEIT International Annual Conference (AEIT), pp. 1–6, October 2015. IEEE (2015)

Cyber-Physical Energy Systems Approach for Engineering Power System State Estimation in Smart Grids Seshadhri Srinivasan1(&), Øystein Hov Holhjem2, Giancarlo Marafioti2, Geir Mathisen2, Alessio Maffei3, Giovanni Palmieri3, Luigi Iannelli3, and Luigi Glielmo3 1

Berkeley Education Alliance for Research in Singapore, Singapore 138602, Singapore [email protected] 2 SINTEF ICT-Applied Cybernetics, Trondheim, Norway {Oystein.Hov.Holhjem,fgiancarlo.marfioti, geir.mathiseng}@sintef.no 3 Department of Engineering, University of Sannio, Benevento, Italy {amaffei,palmieri,luiannel,glielmo}@unisannio.it

Abstract. This investigation proposes a CPES architecture and model for engineering energy management application for smart grids. In particular, the investigation considers the implementation of the power systems state estimator (PSSE). The CPES architecture has three layers: physical, monitoring and applications. The physical layer consists of the grid and the various components. Since, the grid is usually engineered with various devices from multiple vendors that have different protocols and standards; data aggregation becomes a problem. The second layer of the CPES architecture overcomes this problem by proposing a middleware that aggregates data from the physical layer. The topmost layer is the applications layer, where the energy management system applications are implemented. These applications require the model, topology and information from the grid. This requires combining the physical aspects of the grid with the cyber ones. This investigation uses the common information model to model the grid and information exchanges. Then the model is combined with measurement and optimization models of the application to realize the PSSE. The proposed approach is illustrated on a Norwegian distribution grid in Steinkjer. Our results show that the CPES approach provides an easier way to engineer future smart grid applications. Keywords: Cyber-physical energy systems (CPES)
Power systems state estimator (PSSE)
Common information models (CIM)
Service oriented architecture (SOA)
Middleware

1 Introduction Traditionally, core responsibility of power engineers has been to operate power grids and manage the transfer of energy. With the advent of smart grids there is an unprecedented volume of data available that needs to be harmonized with the energy © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_11

Cyber-Physical Energy Systems Approach for Engineering Power System

119

flow. Consequently, communication and energy infrastructure have to be designed together for achieving cost benefits from smart grids. The communication infrastructure includes middleware, communication networks, data-models, and software platforms. Furthermore, since smart grids are massively distributed systems scalability plays a vital role within various applications such as state estimation, stability analysis and contingency analysis. These applications rely on measurements from heterogeneous sensors that are manufactured by various vendors. Engineering such applications requires a framework for handling both the cyber and physical parts of the grid simultaneously. Cyber Physical Energy Systems (CPES) Approach is touted as a futuristic approach for engineering smart grids. They consider the tight coupling between the physical and cyber domains in their design. CPES is a recent research topic that has attracted significant attention [1]. Palensky et al. [2] proposed to study the modern energy systems using the CPES approach with the following categories: physical models, information technology, roles and individual behavior, and aggregated as well as stochastic models. Furthermore, proposed the use of CPES approach as a tool and method for the emerging complexities of the energy grids. While, specialized and useful tools for individual domains of the energy systems exist, there is no methodology to combine the different aspects of the energy and information. Motivated by this CPES approach has been studied for modelling and enabling various aspects of the energy grids. To our best knowledge, Ilic et al. [3] first proposed a dynamic model using the CPES approach. The proposed model greatly dependent on the cyber technologies supporting the physical system. Major contributors to the CPES modelling and design are Widl et al. [4–7]. In [4], the author studied the role of the continuous time and discrete-event CPES models for control applications such as demand response, load balancing, and energy storage management. The investigation [7] presented a co-simulation platform for components, controls, and power systems based on software applications such as (GridLAB-D, OpenModelica, PSAT, 4DIAC) for smart charging of electric vehicles. The use of CPES approach for energy optimization in buildings has been studied in [8–10]. The use of model based design methodology for validating control algorithms in a residential microgrid has been studied in [11]. The role of CPES system for electric charging applications has been studied in [12]. Similarly, reducing energy consumption by combining CPES and industry 4.0 was proposed in [13]. More recently, McCalley et al. [14, 15] surveyed the design techniques and applications of CPES. In spite of these developments, the role of CPES approach for handling grid level applications has been minimum. Furthermore, CPES approach encapsulating both the physical and cyber domain has not been investigated. Among the various problems that can be tackled using CPES approach, we focus on the power system state estimation problem that has been studied in energy grids. The PSSE determines the most likely state of the power system from a set of measurements that are captured using supervisory control and data-acquisition systems. The role of PSSE is crucial for many energy management applications such as optimal power flow, monitoring, and contingency analysis. In addition, PSSE also provide snapshots of the network to energy management system applications developed by many third party applications. The role of PSSE in enabling optimal power flow [16, 33], bad-data detections, market

120

S. Srinivasan et al.

operations etc. [17, 32] has been studied in literature. Currently, there are various tools, standards and models used for capturing the various aspects of PSSE. The physical models used for PSSE have been reviewed in [18]. Among the communication standards used for modelling PSSE, IEC 61850 [19] and Common Information Models (CIM) standardized by the International Electromechanical Commission have been studied for implementing the grid. The investigation in [20] viewed PSSE as one of the applications enabled by IEC 61850 without much description about the implementation. The role of IEC 61850 in implementing PSSE for a distribution grid has been proposed in [21]. Similarly, CIM models for state estimation have been proposed in [22–24]. The recent research are converging more towards the use of CIM as a standard for data communication in smart grids (see, [25–30] and references therein). The main contribution of the paper is a CPES architecture and model that can be used to engineer smart grids. The proposed architecture has three layers: physical, middleware and applications. The physical layer is the grid with the various sensors and devices. Typically the grid is composed of various components and devices that have various communication protocols, standards and manufactured by different vendors. Aggregating information from them is difficult. To overcome this, a middleware is used. The top layer of the CPES architecture is the application layer that needs to obtain the data from the middleware map it into the grid model. To model the grid, we use the common information model as it is easily extendable to model various aspects of the grid. The CIM model is tailored to the needs of the PSSE and is used for information exchange to the applications layer as well as among the applications. The application layer in addition contains the optimization and measurement models for mapping the physical variables of the grid. Furthermore, the implementation aspects of the PSSE in the application layer is also presented. The CIM based model maps the cyber aspects with the network topology and to the physical entities directly. Furthermore, the PSSE is also simultaneously realized in the application layer. This leads to a new approach for modelling and realizing future energy management applications. It is worth to mention here that the middleware part of the architecture is beyond the scope of this investigation and has not been treated in detail. The paper is organized as follows. Section 2 presents the implementation aspects of the PSSE. Section 3 presents the CPES architecture and model. Section 4 presents the case study and implementation in the distribution network in Steinkjer. Conclusions and future course of investigation are discussed in Sect. 5.

2 Power System State Estimation The main objective of a power system state estimator is to estimate the state of each bus, namely voltage magnitude and phase angle, using the set of redundant measurements, usually affected by errors. The set of phasors representing the complex bus voltages are called static state of the system. The State estimator, initially developed as an engineering tool, is slowly transforming into a decision support system that assists control and monitoring tasks. As with any estimator, the SE outcome is an approximation of the current network state. Moreover, from a high level supervisory control and data acquisition module, sensors measurements arrive with different time-granularities so

Cyber-Physical Energy Systems Approach for Engineering Power System

121

that measurements and state estimate can differ in phase. The physical component of the PSSE can be broken down into two parts: hardware and software. The hardware part consists of the SCADA system that collects the real-time information from the remote terminal units installed in various substations. RTU measurements include active and reactive power flows, power injections, current magnitude, voltage magnitude and phase angles. Figure 1 shows the hardware components of PSSE. Till recently, a direct measurement of voltage phase angles was considered impossible. Then, the introduction of phase measurement units (PMUs) has been made that measurement possible. In recent times, the PMUs have significantly improved the measurements accuracy by using the GPS to synchronize the time signals with an accuracy close to 1 µs.

Fig. 1. Hardware components of the PSSE

The software components of the PSSE are: measurement pre-filtering, topology processor, observability analysis, state estimator, bad-data detection and bad-data suppression. The pre-filtering block checks for error measurements and unusual data. The topology processor builds the network topology required for the PSSE application using the knowledge of the physical topology. The SE algorithm then computes the state estimates. Finally, the bad data processor checks for the bad data and eliminates the error values. The state estimator block uses iterative algorithms that have finite convergence time for building the PSSE. In summary, the SE operation serves as a large-scale filter between remote measurements and high-level applications that constitute the Energy Management System (EMS).

3 CPES Model Energy management in smart grids require orchestrating several hardware platforms, protocols and devices from multiple vendors. The EMS applications should get access to different data from these various devices. A good solution to solve the heterogeneity

122

S. Srinivasan et al.

issue is to design a middleware. The architecture of such a middleware is shown in Fig. 2. The bottom most layer is the physical layer that models the energy grid. The middleware aggregates the data from the grid and provides it to the PSSE application. The application layer houses the third party and other applications for EMS such as the PSSE. Energy grids typically consider the energy flows in the physical grids, however to orchestrate the smart grid, the information from heterogeneous sensors needs to be aggregated by the middleware and provided to the applications. The physical layer models the energy flows and transfers. This model needs to be augmented with the data model for implementing the EMS. The network topology and other physical aspects needs to be captured in the application layer. The flow of information among the different layers is defined by the CIM model. Here, we do not consider the implementation aspects of the middleware and the services are not elaborated. To propose the CPES model, we propose a top-down approach wherein the application layer has the PSSE application. The PSSE application requires measurements from the grid that is aggregated using the middleware. To process the information, the topology and model of the grid needs to be known. For this purpose the CIM models are used.

Fig. 2. CPES layered architecture

The PSSE EMS application is developed in JAVA. In order to facilitate information interoperability an application program interface (API) is defined. This API can enable different applications and services to access data and exchange information independently, alleviating difficulties with intrinsic data representation. The CIM specifies the semantics of the API. This also eliminates the difficulty of interfacing lower level

Cyber-Physical Energy Systems Approach for Engineering Power System

123

components with the physical application. The CIM model can be used to build semantics into the middleware as well. The lower level components transmit the information using dedicated communication infrastructure. There are different information models and data formats for the data. As the CIM provides interface to the applications, they are oblivious to the physical entities connected to the grid. This leads to a service oriented architecture for the middleware. In what follows we describe the measurement, optimization and grid model captured using CIM. The implementation aspects of the application layer and different aspects considered in the design of PSSE. A. Measurement Model The State Estimator routine is used to monitor the power network status during normal operation, where the system is in quasi-steady state responding to slowly varying load demand and power generation. It is also assumed that, in the three-phase transmission system, all loads are balanced and all transmission lines are fully transposed so that the network can be represented by its single phase equivalent diagram. Given the single phase diagram, the network is mathematically modelled and all the measurements, previously described Sect. 2, are written as function of the network state variables (i.e. voltage phasors). The equations modelling the power network are nonlinear and do not take into account all possible errors due to the uncertainties in the network parameters, e.g., metering errors and noise that may be introduced through the telecommunication systems. Consider a vector z as the set of all available measurements. This vector can be expressed in terms of power network state variables x as follows: 2

3 2 3 2 3 h1 ðx1 ; x2 ; . . .; xn Þ z1 e1 6 z2 7 6 h2 ðx1 ; x2 ; . . .; xn Þ 7 6 e2 7 6 7 6 7 6 7 z ¼ 6 .. 7 ¼ 6 7 þ 6 .. 7 ¼ hðxÞ þ e .. 4 . 5 4 5 4 . 5 . zm

hn ðx1 ; x2 ; . . .; xn Þ

ð1Þ

em

where h(x) is the measurement model, x is the state vector, and e is the error in measurements. The measurement model h(x) is a nonlinear model that maps the state vector to the measurement. B. Optimization Model The mathematical formulation of the presented state estimation problem is based on the maximum likelihood concept. Maximum likelihood estimator (MLE) of a random variable maximizes the likelihood function, which is defined based on assumptions of the problem formulation. The first assumption, as previously mentioned, is that the errors are distributed according to a Gaussian (or normal) distribution, with the expected value equal to zero. Thus, a random variable z is said to have a normal distribution if its probability density function f ðzÞ is given as follows: 1 1 2 f ðzÞ ¼ pffiffiffiffiffiffiffiffi e2ðzlÞ=r 2pr

ð2Þ

124

S. Srinivasan et al.

where l is the expected value (or mean) of z ¼ EðzÞ and r is the standard deviation of z. The previous property holds for the i.i.d. assumption. The second assumption implies that the joint probability density function (pdf) of a set of m measurements can be obtained by taking the product of individual pdfs corresponding to each measurement. The result product function is called likelihood function for the set of m measurements: fm ðzÞ ¼ f ðz1 Þf ðz2 Þ. . .f ðzm Þ

ð3Þ

Essentially the likelihood function is a measure of the probability of observing a given set of measurements in vector z. For this reason we are interested in finding the parameter vector that maximize this function. In order to simplify the procedure of determining the optimum parameters, the function is commonly replaced by its logarithm, obtaining the so called log-likelihood function. Hence, the MLE of the state x can be found by maximizing the log-likelihood function for a given set of observations, z1 ; z2 ; . . .; zm . Thus, the following optimization problem is formulated: minimize JðxÞ ¼

 m  X zi  l 2 i

ri

i¼1

ð4Þ

Let us define, ri ¼ zi  li is the residual of measurement i. The expected value E ðzi Þ of measurement zi , can be expressed as li ¼ E ðzi Þ ¼ hi ð xÞ, where hi ð xÞ is a nonlinear function relating the system state vector x to the i th measurement. The minimization problem in (4) can be modelled as an optimization problem for the state vector x: min x

s:t:

m X

Wii ri2

ð5Þ

i¼1

zi ¼ hi ðxÞ þ ri

8i ¼ 1; . . .; m

where Wii ¼ r2 is the inverse of the assumed error variance for the measurements. i The reciprocal of the measurement variances can be thought of weights assigned to individual measurements. High values are assigned for accurate measurements with small variance and low weight for measurements with large uncertainties. A little manipulation leads to min

JðxÞ ¼ r T Wr

s:t:

z ¼ hðxÞ þ r

x

ð6Þ

where x; r; z; hð xÞ are vectors of respective quantities and W is the diagonal weighted matrix of the WLS problem equals to the inverse of covariance matrix of the measurements.

Cyber-Physical Energy Systems Approach for Engineering Power System

125

C. Smart Grid Model As stated earlier CIM has been chosen as a representative mechanism for the smart grid model. The CIM models provide an abstract model for the power network using unified modelling language (UML). The CIM represents the power system entities using an object oriented approach as classes, attributes, methods and association as defined in IEC 61970 [25] and IEC 61968 [26] standards. The standard IEC 61970-301 provides as semantic model for the power system components at an electrical level and their interrelationships. The IEC 61968-11 extends the semantic models to include the data exchanges for scheduling, asset management and other market operations. Although, CIM contains most classes and their associations to represent the power system, still object models need to be adapted for implementing specific application. CIM models can be adapted by defining new classes, subclasses, methods and attributes. CIM Profile: A profile is a delimitation of CIM which consists of a subset of classes and attributes that specify information conceptually and the relationships among the different objects. A subset of IEC 61970-456 defines the CIM necessary to describe the PSSE results. A modular approach is used in the development of CIM profile that four profiles: Equipment, Measurement, Topology, and State variable Profile. The equipment profile models physical elements of the network such as network, loads, generators, and switches. The measurement profile contains the measurement information such as active and reactive power, voltages, load angles etc. Topology profile defines the classes needed to describe the network topology considering the switching status. State variable profile contains the model for defining the state variables in the network. The relationship between the different profiles is shown in Fig. 3. The profile connected at the “from” end of the arrow depends on the profile at “to” end of the arrow.

Fig. 3. CIM profiles for implementing the PSSE

126

S. Srinivasan et al.

As a next step the abstract and concrete classes of the various profiles are defined. As the treatment of all the profiles is beyond the scope of this investigation. Here we provide an example of the topology profile (see, Fig. 4). The topology profile defines the classes needed to describe how each of the equipment in the network is connected to each other. Topology is given by the association of the buses with the corresponding association of the terminals of the equipment. This way the network model is built as a branch flow model and can be directly used by the PSSE application. The state variable

Fig. 4. Topology profile

Cyber-Physical Energy Systems Approach for Engineering Power System

127

profile of the CIM model is shown in Fig. 5. Similarly, the other CIM profiles are modelled to describe the network. In addition the abstract and concrete models required of the CIM profiles are described. The CIM model maps the network topology with the measurements and state variables. The CIM model defines the semantics for the PSSE API which is implemented in JAVA. The application uses the middleware services to query the sensors and update the network state using the PSSE application. The middleware services required can be broken down into low-level services and common services. The description of the middleware services are not considered in this investigation. REST interfaces are used to communicate between the middleware and application layer.

Fig. 5. State variable profile

D. Software Model of the PSSE Application The PSSE application needs to map the data aggregated from the middleware to the network topology. To this extent the application uses the InterPSS that defines the network data in the IEEE format for processing. The application implementing the PSSE is written in JAVA and to solve the optimization model in (6) GAMS solver is

128

S. Srinivasan et al.

used due to its symbolic processing capability. The APIs implementing the CIM, and libraries of the InterPSS, GAMS are integrated with the application written in JAVA to realize the PSSE for the EMS. The CPES model described above integrates the physical equations describing the energy flows in the network with the CIM to model the cyber part of the network. The resulting model encapsulates both the cyber and physical aspects of the network. Furthermore, the application software model proposed also uses the InterPSS to map the topology and uses the GAMS for realizing the PSSE. This model along with the CPES description of the network can be used to engineer energy management system application in typical smart grid. The main advantage of the proposed approach is that the method provides a comprehensive framework for co-designing both the physical and cyber aspects of the network, thereby reduces the engineering efforts. Furthermore, provides an interface between the different domain engineers involved in smart grids. The CPES model described above integrates the physical equations describing the energy flows in the network with the CIM to model the cyber part of the network. The resulting model encapsulates both the cyber and physical aspects of the network. Furthermore, the application software model proposed also uses the InterPSS to map the topology and uses the GAMS for realizing the PSSE. This model along with the CPES description of the network can be used to engineer energy management system application in typical smart grid. The main advantage of the proposed approach is that the method provides a comprehensive framework for co-designing both the physical and cyber aspects of the network, thereby reduces the engineering efforts. Furthermore, provides an interface between the different domain engineers involved in smart grids.

4 Case Study The CPES approach proposed in Sect. 3 was used to engineer PSSE in the energy management system of the distribution grid in Steinkjer, Norway. It is a radial distribution network that consists of a: hydro power plant with 2 generators, 32 aggregating loads, 50 link busses, and 84 transmission lines. The CIM defines the basic ontology of the set of attribute value pairs. Ontologies are often written in XML or in a Resource Document Framework (RDF), which is a suitable format for middleware information exchange through the common communication bus. The CIM standard IEC 61970-52 defines the procedure for description of the network model as a serialized RDF schema. The main concept of the RDF is called the tiple and consists of subject-predicate-object expression. An RDF document contains element that are identified by unique ID attribute and that can be referenced from other elements using that ID in a resource attribute. The CIM profile for implementing the PSSE can be represented using a RDF schema document. In details, after having selected all the classes, attributes and relationship among the classes, tools such as CIMTool can be used for generating the RDF schema. As stated earlier the PSSE application uses the InterPSS and GAMS to provide the state estimates for the different applications. The CIM model also defines the information exchange among the EMS applications. The CPES engineered PSSE application is tested in the pilot distribution network. The residuals of the PSSE are used to evaluate the accuracy of the state estimates. Figure 7 shows the residuals

Cyber-Physical Energy Systems Approach for Engineering Power System

129

computed for the real and reactive power in p.u. values. Our results show that the PSSE provided an accurate estimates with error between 2–4% for the 85 buses in the distribution network (Fig. 6).

Fig. 6. Residuals of the real power in PSSE

Fig. 7. Residuals of the reactive power in PSSE

5 Conclusions This investigation proposed a cyber-physical energy systems approach for engineering PSSE in smart grids. The proposed approach modelled both the physical and cyber aspects of the PSSE simultaneously. The physical model was represented by the

130

S. Srinivasan et al.

measurement and optimization model. While, the common information model was used to describe the cyber part of the CPES model. The proposed approach can be used to engineer PSSE applications in smart grids with heterogeneous sensors from various vendors and considering interoperability constraints. Furthermore, the proposed CPES approach can also be used for building a service oriented architecture (SOA) based middleware. The details of the implementation of the middleware are not discussed in the investigation. Extending the CPES approach to model optimal power flow with PSSE is the future course of this investigation. Acknowledgement. This research is supported by the European Union’s Seventh Framework program (FP7 2012-2015) for the ICT based Intelligent management of Integrated RES for the smart optimal operation under the grant agreement n. 318184, project name: I3RES.

References 1. Khaitan, S.K., McCalley, J.D.: Cyber physical system approach for design of power grids: a survey. In: 2013 IEEE Power and Energy Society General Meeting (PES). IEEE (2013) 2. Palensky, P., Widl, E., Elsheikh, A.: Simulating cyber-physical energy systems: challenges, tools and methods. IEEE Trans. Syst. Man Cybern.: Syst. 44(3), 318–326 (2014) 3. Ilic, M.D., Xie, L., Khan, U.A., Moura, J.M.F.: Modeling future cyber-physical energy systems. In: IEEE Power and Energy Society General Meeting-Conversion and Delivery of Electrical Energy in the 21st Century, pp. 1–9. IEEE (2008) 4. Palensky, P., Widl, E., Elsheikh, A.: Simulating cyber-physical energy systems: challenges, tools and methods. IEEE Trans. Syst. Man Cybern.: Syst. 99, 1–10 (2012) 5. Widl, E., Palensky, P., Elsheikh, A.: Evaluation of two approaches for simulating cyber-physical energy systems. In: Proceedings of 38th Annual Conference on Industrial Electronics, IECON 2012, pp. 3582–3587 (2012) 6. Stifter, M., Widl, E., Filip, A., Elsheikh, A., Strasser, T., Palensky, P.: Modelica-enabled rapid prototyping of cyber-physical energy systems. In: IEEE Workshop on Modeling and Simulation of Cyber Physical Energy Systems, pp. 1–6 (2013) 7. Elsheikh, A., Awais, M.U., Widl, E., Palensky, P.: Co-simulation of components, controls, and power systems based on open source software. In: Proceedings of 2013 IEEE Power and Energy Society General Meeting, pp. 1–5 (2013) 8. Kleissl, J., Agarwal, Y.: Cyber-physical energy systems: focus on smart buildings. In: Proceedings of the 47th ACM Design Automation Conference, pp. 749–754 (2010) 9. Zhao, P., Godoy, S.M., Siddharth, S.: A conceptual scheme for cyberphysical systems based energy management in building structures. In: 9th IEEE/IAS International Conference on Industry Applications (INDUSCON), pp. 1–6 (2010) 10. Shein, W.W., Tan, Y., Lin, A.O.: PID controller for temperature control with multiple actuators in cyber-physical home system. In: 15th IEEE International Conference on Network-Based Information Systems (NBiS), pp. 423–428 (2012) 11. Ge, Y., Dong, Y., Zhao, H.: A cyber-physical energy system architecture for electric vehicles charging application. In: 2012 12th International Conference on Quality Software (QSIC). IEEE (2012) 12. Al Faruque, M.A., Ahourai, F.: A model-based design of cyber-physical energy systems. In: Design Automation Conference (ASP-DAC), 2014 19th Asia and South Pacific. IEEE (2014)

Cyber-Physical Energy Systems Approach for Engineering Power System

131

13. Martin, B., et al.: A new approach to increasing energy efficiency by utilizing cyber-physical energy systems. In: 2013 Proceedings of the 11th Workshop on Intelligent Solutions in Embedded Systems (WISES). IEEE (2013) 14. Khaitan, S.K., McCalley, J.D., Liu, C.C. (eds.): Cyber Physical Systems Approach to Smart Electric Power Grid. Springer, Heidelberg (2015) 15. Khaitan, S.K., McCalley, J.D.: Design techniques and applications of cyberphysical systems: a survey. IEEE Syst. J. 9(2), 350–365 (2014) 16. Maffei, A., Srinivasan, S., Iannelli, L., Glielmo, L.: A receding horizon approach for the power flow management with renewable energy and energy storage systems. In: 2015 AEIT International Annual Conference (AEIT), pp. 1–6. IEEE, October 2015 17. Huang, Y.-F., et al.: State estimation in electric power grids: meeting new challenges presented by the requirements of the future grid. Signal Process. Mag. 29(5), 33–43 (2012). IEEE 18. Kashyap, N.: Novel resource-efficient algorithms for state estimation in the future grid. Ph.D. dissertation, Department of Electrical Engineering, Aalto Unieristiy, Finalnd (2012) 19. Srinivasan, S., Kumar, R., Vain, J.: Integration of IEC 61850 and OPC UA for Smart Grid automation. In: 2013 IEEE Innovative Smart Grid Technologies-Asia (ISGT Asia), pp. 1–5. IEEE, November 2013 20. Mohagheghi, S., et al.: Applications of IEC 61850 in distribution automation. In: 2011 IEEE/PES Power Systems Conference and Exposition (PSCE). IEEE (2011) 21. Choi, S., Kim, B., Cokkinides, G.J.: Feasibility study: autonomous state estimation in distribution systems. IEEE Trans. Power Syst. 26(4), 2109–2117 (2011) 22. Xu, K.-N., Xin-Gong, C., Xi-Ji, X., Yan-Shi, C.: Model design of electric system state estimation based on CIM. In: Power and Energy Engineering Conference, 2009. APPEEC 2009. Asia-Pacific. pp. 1–4. IEEE (2009) 23. LiJun, Q., Meng, L., Ying, W., CuiJuan, H., HuaWei, J.: Cimbased three-phase state estimation of distribution network. In: 2011 International Conference on Advanced Power System Automation and Protection (APAP), vol. 1, pp. 667–672. IEEE (2011) 24. Sharma, A., Srivastava, S., Chakrabarti, S.: An extension of common information model for power system multiarea state estimation. Syst. J. PP(99), 1–10 (2014). IEEE 25. Chen, Y.: Industrial information integration—a literature review 2006–2015. J. Ind. Inf. Integr. 2, 30–64 (2016) 26. Cintuglu, M.H., Martin, H., Mohammed, O.A.: An intelligent multi agent framework for active distribution networks based on IEC 61850 and FIPA standards. In: 2015 18th International Conference on Intelligent System Application to Power Systems (ISAP), pp. 1– 6. IEEE, September 2015 27. Belikov, J., Kotta, Ü., Srinivasan, S., Kaldmäe, A., Halturina, K.: On exact feedback linearization of HVAC systems. In: 2013 International Conference on Process Control (PC), pp. 353–358. IEEE, June 2013 28. Soudari, M., Srinivasan, S., Balasubramanian, S., Vain, J., Kotta, U.: Learning based personalized energy management systems for residential buildings. Energy Build. 127, 953–968 (2016) 29. Srinivasan, S., Kotta, U., Ramaswamy, S.: A layered architecture for control functionality implementation in smart grids. In: 2013 10th IEEE International Conference on Networking, Sensing and Control (ICNSC), pp. 100–105. IEEE, April 2013 30. Landa-Torres, I., et al.: The application of the data mining in the integration of RES in the smart grid: consumption and generation forecast in the I3RES project. In: 2015 IEEE 5th International Conference on Power Engineering, Energy and Electrical Drives (POWERENG). IEEE (2015)

132

S. Srinivasan et al.

31. Cavalieri, S., Regalbuto, A.: Integration of IEC 61850 SCL and OPC UA to improve interoperability in Smart Grid environment. Comput. Stand. Interfaces 47, 77–99 (2016) 32. Verrilli, F., et al.: Model predictive control-based optimal operations of district heating system with thermal energy storage and flexible loads. IEEE Trans. Autom. Sci. Eng. 14(2), 547–557 (2017) 33. Maffei, A., Srinivasan, S., Castillejo, P., Martinez, J.F., Iannelli, L., Bjerkan, E., Glielmo, L.: A semantic middleware supported receding horizon optimal power flow in energy grids. IEEE Trans. Ind. Inform. PP(99), 1. doi:10.1109/TII.2017.2655047. http://ieeexplore.ieee. org/stamp/stamp.jsp?tp=&arnumber=7822985&isnumber=4389054

Knowledge-Based Technologies for Web Applications, Cloud Computing, Security Algorithms and Computer Networks

About the Applications of the Similarity of Websites Regarding HTML-Based Webpages Doru Anastasiu Popescu1(&), Ovidiu Domșa2, and Nicolae Bold3 1

Faculty of Mathematics and Computer Science, University of Pitesti, Pitesti, Romania [email protected] 2 “1 Decembrie 1918” University of Alba Iulia, Alba Iulia, Romania [email protected] 3 Faculty of Management, Economic Engineering in Agriculture and Rural Development, University of Agronomic Sciences and Veterinary Medicine Bucharest, Slatina Branch, Bucharest, Romania [email protected]

Abstract. The study of the similarity between web applications has extended alongside with the informational explosion resulted from the fast communication means through Internet. The copyright of web applications is difficult to be appreciated in this domain and this is the reason for the development of novel web technologies and mechanisms of measuring the similarity between two webpages. In this paper, we will present a modality of measurement of the similarity degree between two webpages regarding the HTML tag-based webpages. The degree of similarity will be determined approximately, being dependent of the webpages used from the both websites and the tags set used in the comparison of the webpages. The selection of webpages in order to determine the degree of similarity between two webpages will be made using genetic algorithms. In the final part of the paper there are presented the results obtained with the implementation of the algorithm presented in the paper. Keywords: Webpage


Tag
Algorithm
Chromosome
Gene
Similarity

1 Introduction The similarity is a notion which appears frequently in many domains: science, linguistics, copyright, arts, transportation etc. In the last period, different mechanisms of measuring the similarity between different objects were developed, using particularly complex algorithms and mathematical results. A variant of measuring the similarity between two objects presumes the identification of certain character sequences which characterize adequately these objects and then these sequences are analyzed by means of specific algorithms to obtain a number which reflects the similarity. This type of algorithms is used in [4, 5]. We will use this idea in this paper for measuring the similarity of two webpages. The number that will © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_12

136

D.A. Popescu et al.

be obtained to define to what extent two webpages are similar will be determined by examining the tag sequences which are used at the web pages created using HTML. The large number of files used for creating the webpages is an impediment for creating algorithms which calculate the similarity between two webpages and this is the reason for which we will use a fixed number of files for this calculus. The generation of these files will be made using genetic algorithms, which will provide several variants of these files and for each variant we will use the editing distance method for the calculus of similarity, presented in detail in [7]. The main motivation for our work is the finding of the degree of similarity between two web applications. Another motivation is related to finding different methods of calculating this degree, in this paper being presented a method using genetic algorithms. In Sect. 2 we will present some definitions and examples necessary for the approximate calculus of the degree of similarity between two webpages. Section 3 will contain the presentation of the selecting modality required for the measurement of the similarity between two web pages, using genetic algorithms. In Sect. 4, we will present the modality of calculating the similarity of webpages using the results from Sect. 3. Section 5 is dedicated to the presentation of several results obtained after the algorithms presented in Sects. 3 and 4 were implemented.

2 Preliminary Notions In order to determine the similarity between two webpages which will be denoted by WA (web applications): WA1 and WA2, we will use only the files formed of HTML tags. These files have the .html and .htm extensions. To simplify the future notations, we will denote by P = {p1, p2, …, pm}, respectively with Q = {q1, q2, …, qn} the sets of web pages associated with the applications WA1, respectively WA2. Next, we will interpret by pi or qj the file which contains the tags of the corresponding web page (1  i  m, 1  j  n). In the approximate calculus of the similarity between two web pages we will not take into consideration a fixed set of tags, this set being denoted by TG. For the next definitions, it is useful to denote by T1i the sequence of tags from pi which are not found in TG and by T2j the sequence of tags from qj which are not found in TG, where 1  i  m, 1  j  n. In the calculus of the similarity between two sequences of tags we will use the editing distance (or the Levenshtein distance), presented in detail in [3, 7, 11]. In the classic definition, the character strings are used, in this paper the characters being replaced by tags. In the paper [7] we have done a similar study. We will use, in this matter, some notations used in the referred paper: – DL(s, t) is the minimum number of operations to transform the string s into the string t; – d(pi, qj) is the degree of similarity between two webpages pi and qj;

Similarity of Websites Regarding HTML-Based Webpages

137

– d(pi, WA2) is the degree of similarity between the webpage pi and the web application WA2; – d(WA1, WA2) is the degree of similarity between two web applications WA1 and WA2; – ad(WA1, WA2) is the approximate degree of similarity between two web applications WA1 and WA2.

3 The Choice of Files for the Calculus of the Similarity Between Two Webpages At the notations used in the previous sections we add two natural numbers k and h with 1  k  m, 1  h  n. k and h are chosen so that the runtime of the algorithm of the calculus of the similarity between two web pages to be enclosed in limits closed to the ne calculated for all the web pages. In Fig. 1 the general scheme of the algorithm as in [3, 8] is presented. The algorithm will be used twice for generating subsets with k web pages from WA1, respectively with h web pages from WA2.

Initial Population NoGeneration++ Select next generation

Mutation NoGeneration > NoF

Sort the chromosome

Perform Crossover

Exit Fig. 1. Scheme of a genetic algorithm

The algorithm presented next has as input data a string of characters which represents the path where a web application can be found and a natural number k, less or equal with the number of web pages built using HTML tags, denoted by n. The web

138

D.A. Popescu et al.

pages will be codified with numbers from 1, 2, … to n. The output data of the algorithm will be represented by a bi-dimensional array with k columns, each line representing a chromosome, namely a subset with k distinct elements from {1, 2, …, n}. The subsets obtained from the array will be distinct. The steps of this algorithm are: Step 1. The character sequence s and the number k is read. This sequence represents the path of the web application. Step 2. The files with the extension .html or .htm from the tree with the subfolders that are rooted in s are determined. The names of these files will be memorized alongside their path in the string p[1], p[2], …, p[n]. Step 3. The number of generations NoF which will be used in the genetic algorithm is read. Step 4. Using the genetic algorithm presented above, it is built a bi-dimensional array with the elements Pop[i][j], i = 1, Nocrom and j = 1, k, where Nocrom represents the number of chromosomes which will be obtained. Nocrom can be read. Step 5. The chromosome obtained at step 4 are sorted ascendingly depending on the fitness functions, but these chromosomes are not distinct. This is why a new bi-dimensional array will be built. This array contains the elements PopDist[i][j], i = 1, NoCromDist and j = 1, k. Observations 1. In the genetic algorithm from the step 4, the initial population is randomly generated using the numbers from the set {1, 2, …, n}, these numbers being the genes that form the chromosomes, n being the number of web pages from WA. The chromosome is formed from k genes. The initial population used in the implementation presented in Sect. 6 has 1200 chromosomes. 2. For the ith chromosome from the population which contains the genes Pop[i][1], … Pop[i][j], …, Pop[i][k], the fitness is calculated as being the sum of the degrees of similarity of between the pairs of web pages p[Pop[i][j]] and p[Pop[i][r]], 1  j r  k and it is being denoted by f(i). Thus: f ði Þ ¼

k1;k X

d ðPop½i½ j; Pop½i½r Þ

j¼1;r¼j þ 1

3. The ith chromosome is more performant than the jth chromosome if f(i) < f(j), which means that the web pages from the ith chromosome are more different to each other than the ones from the jth chromosome. 4. The mutation operation is made by identifying a number h (randomly generated) from {1, 2, …, n} which is not found in a chromosome and randomly determining a position poz in the chromosome. After that, the gene from the position poz is replaced with h. 5. For the crossover operation, chromosomes which contain at least one different number are randomly generated. The crossover of the two chromosomes with the indices i and j which have this property will be made in this way:

Similarity of Websites Regarding HTML-Based Webpages

139

– it is built a string of numbers with all the genes from the two arrays, denoted by x = (x1, x2, …, x2k). – the array x is sorted ascending and a string of numbers y = (y1, y2, …, yw) containing the distinct numbers form x. w > k, because of the fact that the two chromosomes has at least one gene (number) different. – two new chromosomes are built from the array y, the first chromosome containing the first k components from y and the second with the latter k components from y.

4 The Approximate Calculus of the Degree of Similarity Between Two Webpages In this section, we will keep the notations from the previous sections. The algorithm presented in the next lines will determine the approximate degree of similarity between two webpages WA1 and WA2, using k web pages form WA1 and h web pages from WA2. The steps of this algorithm are: Step 1. Two strings of characters s and t and the natural numbers k and h are read. These two strings represent the paths where the two webpages WA1 and WA2 can be found. Step 2. The files with the extension .html or.htm from the tree with the folders which are rooted in s are determined. The names of these files will be memorized alongside with their paths in the string p[1], p[2], …, p[m]. Step 3. The files with the extension .html or .htm from the tree with the folders which are rooted in t are determined. The names of these files will be memorized alongside with their paths in the string q[1], q[2], …, q[m]. Step 4. Using the algorithm from the Sect. 4 twice, once for WA1 and once for WA2, we determine two bi-dimensional arrays with chromosomes (web pages indexes): ðPopDist1½i½jÞi¼1;NoCromDist1;j¼1;k for WA1 respectively ðPopDist2½i½jÞi¼1;NoCromDist2;j¼1;h ; for WA2: The applications that contains the web pages with the indexes from the chromosomes from the lines of the previous bi-dimensional arrays will be denoted by (WA1i) i = 1, NoCromDist1 respectively (WA2j) j = 1, NoCromDist2. Step 5. We determine ad(WA1i, WA2j) for i = 1, NoCromDist1and j = 1, NoCromDist2. The average of these numbers will be the number which represents the approximation of the similarity between the webpages WA1 and WA2.

140

D.A. Popescu et al.

Observations 1. The presented algorithm complexity is O(Nocrom∙(max(m, n)2 + Nocrom ∙log (Nocrom))). 2. The algorithm presented in this section and the one from the Sect. 4 can be modified in order for other definitions of the similarity between web pages, as the well presented in [6] or [8] or other modalities of measuring the similarity between the character sequences as the ones from [1, 10, 11] to be used.

5 Implementation The implementation made in Java led to results closed to the ones obtained for the similarity between two webpages using the definition 5 from the Sect. 2, meaning the fact that all the web pages from the both webpages were included. Table 1 presents some information about the webpages used at the test of the Java program which implements the algorithms from Sects. 5 and 6. The set of tags used at the implementation of the program was made from the tags:





Table 2 presents some results obtained for the webpages presented in Table 1: the degree of similarity (using the algorithm from [7]) and the approximation of the degree of similarity (using the algorithm presented in this paper) (Fig. 2). Table 1. Webpages using for Java program Notation Number of files Number of HTML files E1 125 23 E2 85 15 E3 90 16 E4 104 24 E5 97 17 E6 2431 476 E7 344 35

The results obtained using the implementation of the algorithms presented in this paper shows the fact that they can be used to calculate the degree of similarity between two webpages with quite an acceptable error (generally, under 0.01). An important aspect is the fact that the presented algorithms can also be used for other modalities of defining the degree of similarity between two files.

Similarity of Websites Regarding HTML-Based Webpages

141

Table 2. Results to implement algorithms in the Java language k 12 12 12 12 15 15 100

WA1 h E1 10 E1 10 E1 15 E1 10 E1 100 E1 20 E6 20

WA2 E2 E3 E4 E5 E6 E7 E7

Similarity degree 0.6527269615774745 0.6528103923698071 0.6390043190028488 0.636510512472533 0.34485409963776364 0.319443831408161 0.568263602360978

Similarity degree

Approximate similarity degree 0.6151740849736574 0.5965780161873493 0.6247214290536568 0.6347343901389451 0.318678530256979 0.291547650419385 0.536853457204536

Approximate similarity degree

0.7 0.65 0.6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 0.2 E1 vs E2 E1 vs E3 E1 vs E4 E1 vs E5 E1 vs E6 E1 vs E7 E6 vs E7 Fig. 2. Results to implement algorithms in the Java language (similarity degree versus approximate similarity degree)

6 Conclusions This paper showed a model for the approximate calculus of the degree of similarity between two webpages, each of them being made from several other components. Particularizing, the applications are webpages and the components are web pages with the source code built from tags. A central thing of this model is the genetic algorithm, which helps us at the generation of component subsets with certain restrictions. This type of algorithms is used in many situations, in [2, 3, 9] being presented some applications of these algorithms. The presented model can also be adapted for other modalities of defining the similarity between the files that compose a web application. For a future work, we propose to create a tool which will calculate the approximate similarity between two webpages using several formulas to calculate the similarities between the files.

142

D.A. Popescu et al.

References 1. Okundaye, B., Ewert, S., Sanders, I.: Determining image similarity from pattern matching of abstract syntax trees of tree picture grammars. PRASA Johannesburg, pp. 83–90 (2013) 2. Caldas, L.G., Norford, L.K.: A design optimization tool based on a genetic algorithm. In: Automation in Construction, ACADIA 1999, vol. 11, no. 2, pp. 173–184 (2002) 3. Darby, S., Mortimer-Jones, T.V., Johnston, R.L., Roberts, C.: Theoretical study of Cu–Au nanoalloy clusters using a genetic algorithm. J. Chem. Phys. 116(4), 1536 (2002) 4. Kim, H.-S., Cho, S.-B.: Application of interactive genetic algorithm to fashion design. Eng. Appl. Artif. Intell. 13(6), 635–644 (2000) 5. Remani, N.V.J.M., Rachakonda, S.R., Kurra, R.S.R.: Similarity of inference face matching on angle oriented face recognition. J. Inf. Eng. Appl. 1(1) (2011) 6. Popescu, D.A., Maria, D.C.: Similarity measurement of web sites using sink web pages. In: 34th International Conference on Telecommunications and Signal Processing, TSP 2011, 18–20 August, Budapest, Hungary, pp. 24–26. IEEE Xplore (2011) 7. Popescu, D.A., Nicolae, D.: Determining the similarity of two web applications using the edit distance. In: 6th International Workshop on Soft Computing Applications, 24–26 July 2014, Timisoara, Romania, 6th IEEE SOFA. LNCS, pp. 681–690 (2014) 8. Popescu, D.A., Dan, R.: Approximately similarity measurement of web sites. In: ICONIP, Neural Information Processing. Proceedings LNCS. Springer, 9–12 November 2015 9. Rahmani, S., Mousavi, S.M., Kamali, M.J.: Modeling of road-traffic noise with the use of genetic algorithm. Appl. Soft Comput. 11(1), 1008–1013 (2011) 10. Dietterich, T.G.: An experimental comparison of three methods for constructing ensembles of decision trees: bagging, boosting, and randomization. Mach. Learn. 40(2), 139–157 (2000) 11. Cormen, T.H., Leiserson, C.E., Rivest, R.R.: Introduction to Algorithms. MIT Press, Cambridge (1990)

Energy Efficient Cache Node Placement Using Genetic Algorithm with Probabilistic Delta Consistency Using Flexible Combination of Push and Pull Algorithm for Wireless Sensor Networks Juhi R. Srivastava(&) and T.S.B. Sudarshan Department of Computer Science and Engineering, Amrita School of Engineering, Bengaluru Campus, Amrita Vishwa Vidyapeetham, Amrita University, Kasavanahalli, Carmelaram P.O., Bangalore 560 035, Karnataka, India [email protected], [email protected]

Abstract. Minimum energy consumption and minimum service delay with maximizing quality of service is key to a worthy WSN. Network lifetime and network connectivity lead to optimization. However, to handle the uncertainties present in WSNs, we need powerful heuristics to solve the optimization problem. A possible way to minimize power consumption is by the use of caching popular data by optimally selecting and placing cache nodes in the network. We propose the use of a powerful searching tool and a well-known soft computing technique; a multi-objective genetic algorithm to achieve optimization goals in the presented work. The proposed work proceeds in two steps to give a complete optimized network system with increased network lifetime. In the first step we use genetic algorithm to carefully select and place cache nodes in the network with aims of maximizing field coverage and minimizing network energy usage. In the second step we perform cache consistency on the cache nodes for valid data retrieval. We use the Probabilistic Delta Consistency (PDC) with Flexible Combination of Push and Pull (FCPP) algorithm. The cache consistency algorithm is implemented on the MAC layer and comparisons are made with 802.11 MAC layer without cache consistency using metrics; routing load, packet delivery ratio, end-to-end delay, normalized load and energy consumed. The network overhead is reduced considerably leading to speedy data access. The algorithm is designed for a clustered environment and is an extension of our previous work where we have successfully proved that genetic algorithm gives better results for node deployment when compared to the state of the art node placement algorithm ScaPCICC. The results are obtained by running the experiments on Matlab and ns2. Keywords: Genetic algorithm
Cache consistency Network
Optimization
Clustering




Wireless Sensor

© Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_13

144

J.R. Srivastava and T.S.B. Sudarshan

1 Introduction A Wireless Sensor Network (WSN) is comprised of a set of several identical sensor nodes with limited CPU and storage capacity and is a special class of ad hoc networks. In a typical example, nodes are capable of sensing environmental attributes, such as temperature, light and humidity, allowing researchers to monitor an area of interest remotely, inconspicuously and continuously. Some of the unique characteristics of a Wireless Sensor Networks like ability to withstand harsh environmental conditions, ability to cope with node failures, mobility, dynamic network topology, communication failures, heterogeneity of nodes, large scale of deployment, unattended operation makes their use in certain applications indispensable. Many applications for WSN have been discussed in the literature. A few include environmental/habitat monitoring, acoustic detection, seismic detection, military surveillance, inventory tracking, medical monitoring, smart spaces, process monitoring and traffic management. However, to work seamlessly and achieve application and user level satisfaction, a WSN faces many challenges. The goal of achieving application-level QoS, with minimum energy consumption and minimum service delay in WSNs becomes a very challenging task. Minimum energy consumption and minimum service delay with maximizing quality of service is key to a worthy WSN. A sensor node failure due to energy depletion would lead to network partitions, path breaks, new route setup and therefore, latency in service and data provisioning. It is observed that most of the energy spent by a sensor node is during listening to the channel, data routing and sensing. Routing or data transmission is a foremost reason for high energy dissipation by sensor nodes. When compared; the amount of energy spent during computation with the energy spent during communication, it is observed that transmitting a single bit through a transceiver expends more energy than computing a single instruction on a processor. Routing in WSNs is data centric than address centric. All sensors detecting and sensing the event report to the sink generating redundancy of messages. However, one disadvantage associated with routing is every time an information is required, it has to be fetched from the source which might be several hops away from the sink. This also invokes participation of all the sensor nodes present in the path thus leading to energy drainage. Literature show tremendous contributions in the design of exceedingly competent routing algorithms for WSNs, however, even with a highly energy efficient data routing protocol, a WSN that is designed for continuous monitoring application can achieve more energy efficiency by other means. Battery life time of a sensor node can be extended if amount of communication and computation done by a node is reduced. A possible way to minimize power consumption is by the use of caching the data. Caching of popular data can save a sensor network from exploitation of many of its scarce resources. To mention a few, caching helps in reducing amount of communication between nodes in the network, reduces network wide transmission, hence, reduces interference because of nodes, overcomes variable channel conditions, reduces network congestion, speeds data access by nodes and provides QoS with minimum energy consumption. However, caching can also considerably impact the system energy expenditure if not implemented efficiently; there are two vital design issues to be addressed in caching; first selection of appropriate nodes to cache data and their

Energy Efficient Cache Node Placement Using Genetic Algorithm

145

optimal positioning, i.e., optimal cache node placement and second, is cache consistency. Because of the mobility associated with nodes, caching strategy also requires cache consistency schemes to ensure that the cached objects are consistent with those stored in the server. Our study focuses on the data access algorithms using cache for WSNs. In literature, many algorithms have been proposed to address these issues to the best possible way but very few work reflect the use of any soft computing techniques to achieve the same. Soft Computing is the fusion of methodologies that were designed to model and enable solutions to real world problems, which are not modeled, or too difficult to model, mathematically. These problems are typically associated with fuzzy, complex, and dynamical systems, with uncertain parameters. The uncertainty associated with the network state information is inevitable in terms of, both, fuzziness and randomness. A powerful advantage of soft computing is the complementary nature of the techniques. Used together they can produce solutions to problems that are too complex or inherently noisy to tackle with conventional mathematical methods. In our work, we have used genetic algorithm as an optimization technique to place the Caching Nodes (CN) optimally. We also perform a Flexible Combination of Push and Pull (FCPP) [4] algorithm on the cached nodes to provide user-specified consistency prerequisites using the Probabilistic Delta Consistency (PDC) model. We have used a clustered WSN network environment with cluster head and cluster members to implement our algorithms. We use MATLAB and ns2 simulators for performing our experiments.

2 Related Works 2.1

Genetic Algorithm

Optimization helps to enhance system performance by selecting optimal point within a given set of strictures or circumstances. The key concept in optimization is to be able to make assessment of the system resources and accordingly make decisions that can minimize the exploitation of these resources to achieve the desired objective with maximized desired output. Considering WSNs, we understand that battery life time of a sensor node can be extended if amount of communication and computation done by a node is reduced. In other words, network lifetime and network connectivity lead to optimization. However, to handle the uncertainties present in WSNs, due distance between nodes, communication channel, we need powerful heuristics to solve the optimization problem and therefore soft computing has been involved to achieve the desired optimization goals in the presented work. The aim of the presented work is to optimize the placement of cache nodes in WSNs. We try to select finest number of nodes in a given WSN to serve as cache nodes for complete network coverage. This optimal number of nodes to be selected as cache nodes is a tricky task as selecting large number of CNs may create redundant messages, caching new information in their small memories will require more enhanced techniques of cache replacement. All this would lead to depletion of vital resources like sensor and network energy, overload network bandwidth and also create network

146

J.R. Srivastava and T.S.B. Sudarshan

congestion. While a smaller number of CNs may not be able to cover the entire application and may lead to network partitions hence leaving areas un-sensed while reporting an event. It has been observed that for many applications, the search space for optimization is excessively large or excessively expensive in terms of computation leading to a NP-hard search space. In such instances it is advised to use random search algorithms that also form the basis for studying non-linear programming techniques in research. Algorithms for optimization under this category are robust to noise, non-linearities and discontinuities. These random search algorithms can adapt to new environmental situations and yield optimized values even when the system parameters cannot be predicted at priori. Random search algorithms are also called as multi-objective programming methods as they can address more than one objective function for a given problem statement while maintaining coherency with the boundary conditions. Algorithms that mimic human brain and behavior have revolutionized the current available optimization methods. Some of these well-known algorithms are Memetic Algorithm (MA), Genetic Algorithm (GA), Shuffled Frog Leaping (SFL), Particle Swarm Optimization (PSO), Ant Colony Systems and Simulated Annealing (SA) [2]. All these techniques try to solve a problem in a comparable way; however, they still differ marginally in employing their optimization procedure. Each of these algorithms employs a depiction to suit the technique and then arrive at an optimal solution iteratively. Genetic Algorithms are powerful techniques in solving optimization problems with huge search spaces with large number of variables where usage of procedural algorithms is complicated. They can solve optimization problems which are non-linear, stochastic, non-differentiable, discontinuous or combinatorial in nature. Genetic algorithms employ evolution process observed in biological populations for selection of best off springs that lead to optimal solutions in successive iterations. This soft computing method generates chromosomes which are the optimization parameters represented in the form of bit strings that allows encoding of both discrete and continuous parameters contained in the same problem statement. An advantage of using GA is that the variables in GA take values within their described constraints and boundaries and therefore, no solutions fall out of the upper and lower limits defined in the algorithm. If at all leak or violations occur it is due to functional restrictions of the design [2]. Authors Khanna et al. [5] have addressed creation of clustering with optimal cluster heads and cluster members in WSNs using GA. Their work is limited in identifying limits in large networks thereby applying GA and hence finding shortest convergence time becomes challenging. In his work, author Yang [11] uses GA for placing sink nodes optimally. Their algorithm helps in fast data delivery without much use of energy during data transmission. However, a major limitation of the work is that sink nodes can be located only at certain sites known as feasible sites. Author Youssef et al. [13], has used GA for energy efficient gateway positioning to reduce data access delay. However, their work is not scalable and therefore does not allow large number of gateways in the network. Authors Yong et al. [12], have used GA in a cluster based WSN to find optimal transmission power of cluster heads and cluster members for data transmission with

Energy Efficient Cache Node Placement Using Genetic Algorithm

147

reduced energy consumption. Their work does not consider nodes which are out of range for transmissions. Authors Mollanejad et al. [6], focus on repositioning of base stations using GA for faster data transmission. They consider a small error bound along with network parameters like residual energy and distance for optimized location allotment. However, they consider a static network with each sensor having a GPS system thereby increasing the cost. Wang et al. [10] have used GA for placing sensors which lead to an optimal path planning. However, they have not considered energy parameter in their algorithm. Author Hoyingcharoen et al. [3], have used GA for selecting optimal number of sensors to guarantee minimum detection probability. However, his work does not consider energy and connectivity of nodes. Sun et al. [9], use GA for deploying relay nodes for complete network coverage. They have modeled their system as a graph and use only a single objective function with a penalty constraint that removes any leaks in the solution from the constraint. The proposed work is an extension of our previous work [7], where we use genetic algorithm against state of the art algorithm ScaPCICC for optimal cache node placement in WSNs. We have been successful in proving that genetic algorithm performs better in terms of reducing data access delay and number of messages in the network, while achieving energy efficient optimal cache node deployment in WSNs. 2.2

Cache Consistency

It is crucial that the data fetched from a cache is not old and represents a fresh version. We can achieve this by using cache consistency schemes that maintain the cached data consistent with those existing in the server (Fig. 1). Cache updates can be initiated by the server, CN or can be a cooperative process. In the server based approach also called push method, invalidation messages are broadcasted by the server to all the CNs

Fig. 1. Consistency models

148

J.R. Srivastava and T.S.B. Sudarshan

indicating the update status of data items; whereas in the second case also known as pull mechanism, it is the CN that polls the server to determine the version of the cached item. However, for some applications a combination of push-pull method can be used. This cooperative cache consistency scheme allows for fresh retrievals of cached data at all points of time, however, cooperative schemes combine the advantages and the disadvantages of both the methods. Depending on the application, the server may or may not keep the record of CNs and their associated data items. If records are maintained then it is called the Stateful (SF) approach otherwise it is known as the Stateless (SL) system. Though SF approach is not scalable but yet it is more efficient as server performs multicast of invalidation reports thus reducing on the number of messages in the network. SL keeps track of the Table 1. Summary of cache consistency schemes Cache consistency schemes Push model

Consistency level

Cache status maintenance

Limitations

STATEFUL

TTR mechanism

Strong consistency Weak consistency Strong consistency Strong consistency

TTL mechanism

Strong consistency

STATELESS

Client Poll/Pull each read mechanism Hybrid mechanism Server invalidation mechanism

Strong consistency

STATELESS

Strong consistency Strong consistency

STATELESS

Overhead caused as server has to maintain state of all cache nodes Since cache initiated pull, cache node miss updated data If lease period is more, critical data might not be updated by server TTR value has to be adaptive based on hot and cold data items in order to maintain stringent consistency requirements No assurance from the server that the datum will not be modified within TTL expires Every read at the cache node adds to round trip delay & network overload due to the generated control messages Disadvantages of both TTL and client poll mechanisms State has to be maintained indefinitely causing wastage of space & many invalidations have to be send when an object is modified causing network overload Long query latency, nodes having same copies query server separately thus increasing overheads & clients in long dose mode may miss IRs Overhead in network caused due to UIR and IR messages floated

Pull model LEASE

STATELESS STATEFUL STATELESS

STATEFUL

IR based invalidation mechanism

Strong consistency

STATELESS

UIR based invalidation mechanism

Strong consistency

STATELESS

Energy Efficient Cache Node Placement Using Genetic Algorithm

149

update history and occasionally floods the updates through invalidation messages generating extensive communication overhead. However, according to application constraints the consistency requirements can be stringent (strong consistency) leading to all time fresh version delivery otherwise can tolerate data which does not guarantee consistency also called weak consistency. If the tolerance of serving a data is never more than more than D time with the data at the server, then it satisfies delta consistency. In MANETS a more common technique used for consistency is called the probabilistic consistency model. Here a query is answered when the data item received back is consistent with the server with a probability of at least C, where C; (0  C  1) is a design pre-defined consistency level. Table 1 gives a summary of various cache consistency schemes [14]. We employ Flexible Combination Of Push and Pull (FCPP) algorithm for cache consistency [4]. The main idea behind using FCPP is that it allows the data source node to delay the update till all the ACKs are received or until the time-out values for all the silent caching nodes expire. FCPP is a strong consistency algorithm that considers Lease protocol as a special case. FCPP benefits users by providing user defined consistency requirements, thus allowing invocation of Pull or Push mechanism whenever necessary. We now proceed to discuss the presented work in detail.

3 Proposed Algorithm 3.1

Cache Node Placement Using Genetic Algorithm - Methodology and Network Model

We propose a multi-objective GA based node placement algorithm for WSNs which optimizes the placement of cache node in WSN by optimizing the network functional parameters like field coverage and number of sensors per CN, while minimizing network constraints like network energy usage, the number of out of coverage or unconnected sensor nodes and number of overlapping CNs. The technique is designed for a clustered WSN used for monitoring applications and focuses on positioning of three types of sensor nodes in a 2-dimensional L  L square grid configuration with predefined Euclidean distance between the units. It identifies nodes as CNs, n1 (high signal range, HSR), n2 (low signal range; LSR). The algorithm also considers inactive sensor nodes which have no role and are silent though not dead in the network. GA in the proposed algorithm tries to situate these nodes optimally so that the field coverage and number of sensors per CN is maximized. The design assumes placing of each sensing node on the intersection points of the square grid for complete network coverage (Fig. 2). The nodes categorized as CN, LSR, HSR, and inactive nodes sensor nodes are encoded in a row by row style with each individual represented by a bit string of size 2L2 (Fig. 3) [7]. For cache node discovery we use Scaled Power Community Index Cooperative Caching (ScaPCICC) [1]. For a given graph G = (V, E), it is defined as

150

J.R. Srivastava and T.S.B. Sudarshan

Fig. 2. Network diagram

Fig. 3. Node representation in WSN

ScaPCICC ¼ PCIðvÞ=CðvÞ; where, PCI(v) = C(v) =

ð1Þ

power community index clustering coefficient

PCI is defined as one hop direct connection or neighbors of node v which is represented as vertex in the graph. Also known as degree of node v and C(v) is described as

Energy Efficient Cache Node Placement Using Genetic Algorithm

cðvÞ ¼ ð2 * LvÞ=ðdv * ðdv  1ÞÞ

151

ð2Þ

where, dv = Degree of node v Lv = # of links among dv neighbor of node v In other words, PCI value of a node indicates its control over other nodes whereas, Scaled PCI gives the minimum cardinality of all the nodes connected in the subgraph and the adjacent nodes in the dominating set of graph G. ScaPCICC displays high accuracy in selection of nodes and can work well for isolated nodes. Nodes selected using ScaPCI become the cache nodes which along with other nodes are encoded in bit string to form chromosomes that can proceed to the next level of optimization process using GA. In the proposed algorithm we use a weighted sum approach to organize all the desired objectives in a single fitness function. F ¼ min

X5



kx i¼1 i i

ð3Þ

where, F is defined as F ¼ a1 FC þ a2 SpCN þ a3 NetE þ a4 OpCnE  a5 SORE

ð4Þ

and values of coefficients ai = 1, 2, 3, .. are decided based on design constraints and experimentations. We define field coverage consisting of components like Ncn (number of CNs sensors), number of nodes with HSR value; Nn1, number of LSR nodes; Nn2, number of inactive sensors; Ninact, number of out of range sensors; Nor and total number of sensing points; Ntotal. We define SpCN as Sensors-per-Caching Node, NetE as network energy, OpCnE is defined as Overlaps-per-Caching node- Error and SORE as Sensors-Out-of-Range Error [7]. 3.2

Cache Consistency Using Flexible Combination of Push and Pull (FCPP)

Once cached nodes are created and popular or most frequently accessed data are cached then the next step is to maintain cache consistency. Cache consistency is necessary to ensure valid data access between the source node that detects the data first and its cached copies. Maintaining cache consistency incurs heavy overhead in WSNs as different users have different consistency requirement. For example some applications require frequent updates like stock market whereas; few applications allow some tolerance level in accessing the data like weather forecast. It therefore, becomes essential to allow users to tune in to their consistency requirements leading to a flexible adaptation to data update rate in the cached nodes. This also results in a trade-off between maintaining user-defined consistency requirements and consistency maintenance cost. The consistency model selected in the presented work is called the Probabilistic Delta Consistency (PDC) method. To further maintain user- specified consistency condition

152

J.R. Srivastava and T.S.B. Sudarshan

at minimum cost we use the Flexible Combination of Push and Pull (FCPP) algorithm. We will discuss each technique in detail. Probabilistic Delta Consistency allows users the flexibility to define their consistency requirements in two dimensions orthogonal to each other. The x-axis indicates the maximum tolerable deviation in d time or value that can exist between the source data and the cached copies. Whereas, the y-axis denotes the probability p, which accounts for the minimum ratio of queries answered by consistent cached copies. PDC model is versatile in setting the consistency requirements in a network. PDC represented as (p, d) helps identify the consistency level of a system. For example PDC (0, 100) indicates strong consistency whereas PDC (*, 0) indicates weak consistency (Fig. 4) [4].

Fig. 4. Probabilistic delta consistency representation

To satisfy the consistency requirement with minimum cost we use the FCPP algorithm. FCPP associates a time-out value with every cached copy based on the d and p values. Cache updates are made by the data source node by sending an (INV) message for which it receives an acknowledgement (INV_ACK) from cached nodes for valid time-out values of the cached copies. However, data updates can be postponed until it receives all the ACKs or until the time-out values of the entire un-responding nodes end or the upper limit of the acceptable delay of data update is reached. The key issue in implementing FCPP is to be able to quantify the tradeoff between consistency and cost. FCPP converts to Pull option when the time-out value 1 is shifts to zero, whereas, it transforms to Push format. FCPP works on both cache node and the data source node. On cache node the algorithm is described below: FCPP on caching node Step 1: Receive data request query (a) IF (l > 0) serve the request with the consistent cache copy; (b) ELSE // the time-out value l is decreased to zero

Energy Efficient Cache Node Placement Using Genetic Algorithm

153

Step 2: generate a RENEW message for a new time-out value and update copy if needed; Step3: Time-out value l is renewed; serve the user request with the updated cache copy; FCPP on the data source node Case 1: The source data is ready to be updated Step1: Send invalidation message (INV) to caching node with positive l; Step2: IF (INV_ACK is received from all cached nodes or D seconds have expired) (2.1) Update the source data; Case 2: On arrival of a RENEW message Step3: IF (source data update is complete) (3.1) Send updates to the caching node; Step4: IF (no pending updates) (4.1) Allow cache nodes to cache copies with time-out values 1; FCPP is implemented on to the MAC layer of the cache node. We move to the simulations and results to see the advantages of the proposed algorithms.

4 Simulation and Results 4.1

GA Simulation and Results

GA Simulation parameters are provided in (Table 2). At the start of the GA simulations, each coefficient were assigned a unity value, however, with course of the experimentation the optimal values of the variables were determined. Figure 5 shows

Fig. 5. Fitness values for the best individuals against average fitness value of the population

154

J.R. Srivastava and T.S.B. Sudarshan Table 2. GA simulation parameters

Parameter Area Crossover Mutation Selection Number of GA runs Number of nodes Stability seen (termination criteria – no change in value) Base station location

Value Grid size of 700 cm  1200 cm Two-point (p = 0.8) Nonlinear (p = 0.1). elitism Roulette wheel 3000 100 (results shown for 34 nodes) 300 runs Node numbered 0, placed at the centre

the fitness values for the best individuals against average fitness value of the population during the GA run. The value stabilizes to 37.8 from 300 runs improving the fitness value to 9.2% at 300 runs and 5.7% at 3000 generations (Fig. 5). In spite of initial randomness due to search for best individuals, the network energy stabilizes to optimum value of 1.73 (Fig. 6).

Fig. 6. Network Energy (NetE)

Figure 7 shows that the OpCnE parameter which is the number of overlaps between the cache nodes successfully reduces as the GA runs proceed towards stabilization. The OpCnE value becomes zero as the experiment progresses from 300 GA iterations.

Energy Efficient Cache Node Placement Using Genetic Algorithm

155

Fig. 7. Overlaps-per-Caching node-Error (OpCnE)

From Fig. 8, it is evident that after the experiment stabilizes the sensors out of range for a cache node is reduced by 11.2% (11.2 mark) and finally diminishes to zero from 300 generations onwards.

Fig. 8. Sensor Out of Range Error (SORE)

156

J.R. Srivastava and T.S.B. Sudarshan

Fig. 9. Sensor-per-Cache Node (SpCN)

Fig. 10. Field coverage (FC)

Figure 9, is aimed at optimizing the fitness function parameter SpCN. The number of Sensors- per- Cache Node graph indicates an increased value of 22 from its previous value of 3.6 mark. This experiment leads to graph in Fig. 10 which shows 89% increase in the filed coverage. Maximizing filed coverage and reduced data access delay with minimum energy usage achieved through optimal cache node selection and placement is the key concept in the presented work. The optimized network obtained from the MATLAB simulator after running GA gives a text file with location information of all the nodes in the WSN. The cluster

Energy Efficient Cache Node Placement Using Genetic Algorithm

157

Fig. 11. Connectivity graph of the optimized network obtained after running GA

heads, cluster members and the identified cache nodes locations coordinates are now fed into the ns2 simulator for maintaining cache consistency requirements. Table 3 shows the text file which is the optimized network obtained after using GA for cache placement in WSN for PCI > 70. The PCI value helps to decide the optimal number of CNs required to cover the entire network. Figure 11 shows the connectivity graph of the optimized network (degree of each node) obtained after running GA over WSN for optimal selection and placement of CNs. From Fig. 11 we can understand the roles of each sensor node in the network for one complete cycle of data transmission. A useful cluster configuration is obtained only if we have a good CH (Cluster head) to CM (Cluster member) ratio so that, clustering covers the entire network using minimum network energy enhancing network lifetime. The optimal ratio of CH to CM for N nodes organized in a clustered architecture is given by CH fraction [8]. CHFrac ¼ CH=N

ð5Þ

where, N is total number of nodes given by following equation N ¼ jCHj þ jCMj

4.2

ð6Þ

FCPP Simulation and Results

FCPP simulation parameters are provided in (Table 4). In the presented, work Flexible Combination of Push and Pull (FCPP) algorithm has been deployed on to the MAC Layer. Comparison of FCPP implemented network with 802.11 MAC layer is done in terms of Packet Delivery Ratio, End-to-end delay, normalized routing load, Routing overhead (in terms of number of packets) and Energy consumed (Tables 5 and 6).

158

J.R. Srivastava and T.S.B. Sudarshan Table 3. Optimized network generated by using the genetic algorithm Node ID X coordinate 1 978 2 604 3 41 4 605 5 571 6 618 7 1065 8 1042 9 36 10 882 11 465 12 47 13 728 14 585 15 331 16 219 17 88 18 1040 19 534 20 811 21 215 22 777 23 543 24 1065 25 796 26 689 27 160 28 941 29 452 30 338 31 1046 32 94 33 255

Y coordinate 437 145 301 337 272 655 614 72 268 464 692 126 275 515 154 633 436 586 621 401 408 664 493 311 388 85 209 74 65 524 636 694 373

Role Head Head Head Member Head Member Member Head Member Head Head Cache node Head Cache node Member Head Head Member Member Cache node Member Head Member Cache node Head Head Member Member Member Cache node Head Member Member

Figures 12, 13, 14, 15 and 16 shows that FCPP performs better over 802.11 MAC layer. Tables 5 and 6 shows a comparative analysis between FCPP implemented MAC layer with 802.11 MAC layer.

Energy Efficient Cache Node Placement Using Genetic Algorithm

159

Table 4. FCPP simulation parameters Parameter Number of nodes Area Base station/sink Maximum node velocity (Vmax) Minimum node velocity (Vmin) Pause time (Tp) Simulation time MAC Consistency requirement p Consistency requirement d Routing protocol Maximum delay before data update Average update interval Mobility model

Value 30–100 700 cm  1200 cm Center of the area (node ID = 0), assumed stationary 20 m/s 0 m/s 0 m/s 100 s IEEE 802.11 60–90% 2 s–20 s AODV 2s 20 s–80 s Random way point

Fig. 12. Packet delivery ratio 802.11 MAC layer versus FCPP implemented MAC layer

160

J.R. Srivastava and T.S.B. Sudarshan

Fig. 13. End-to-end delay 802.11 MAC layer versus FCPP implemented MAC layer

Fig. 14. Normalized routing load 802.11 MAC layer versus FCPP implemented MAC layer

Energy Efficient Cache Node Placement Using Genetic Algorithm

161

Fig. 15. Routing overhead 802.11 MAC layer versus FCPP implemented MAC layer

Fig. 16. Energy consumption 802.11 MAC layer versus FCPP implemented MAC layer

162

J.R. Srivastava and T.S.B. Sudarshan Table 5. 802.11 MAC layer

For pause time (in seconds) 5 6 7 8 9

Packet delivery ratio 60.48 63.56 64.48 59.05 58.49

End-to-end delay

Normalized routing load

Routing overhead

Energy consumed

1.234 1.260 1.36 1.42 1.47

0.954 0.757 0.72 0.64 0.611

289 289 327 301 322

54.48 65.05 82.41 85.29 97.42

Table 6. FCPP implemented MAC layer For pause time (in seconds) 5 6 7 8 9

Packet delivery ratio 81.64 79.37 78.46 77.90 76.69

End-to-end delay

Normalized routing load

Routing overhead

Energy consumed

1.06 1.10 1.13 1.16 1.18

0.174 0.149 0.129 0.114 0.103

71 71 71 71 71

55.53 65.32 75.62 85.89 95.55

5 Conclusions and Future Work In the presented work we propose a scheme for cache node selection and optimal placement of selected cache nodes in WSNs using genetic algorithm. The proposed algorithm uses a multi objective fitness function and aims at minimizing the network energy usage during data transmission, thus resulting in an increased network lifetime. The technique is designed for a clustered WSN used for monitoring applications. The work further progresses by performing cache consistency on the cached nodes to guarantee valid data retrieval by users. To achieve this we use the Flexible Combination of Push and Pull algorithm which is implemented over the 802.11 MAC layer. The algorithm FCPP is run on cached nodes and data source node for valid data access by users. The cache consistency algorithm is implemented on the MAC layer and comparisons are made with 802.11 MAC layer without cache consistency using metrics like routing load, packet delivery ratio, end-to-end delay, normalized load and energy consumed. The results obtained clearly show that cache placement using GA is effective in achieving maximum field coverage of 89% in WSN while reducing network energy consumption thus granting an increase in network lifetime. The cache consistency algorithm FCPP further helps the network in faster data retrieval without message overheads generated in the network. It proves to be a very convenient choice for users to tune their consistency requirements with ease and at the same time avoid too many transmissions in the network thus leading to an energy efficient network system. However, the algorithm FCPP is limited when frequency of data updates are high and the network is unstable. FCPP algorithm performs best when

Energy Efficient Cache Node Placement Using Genetic Algorithm

163

the user requirements are more severe. The experiments were completed using MATLAB and ns2 simulator. For future work, we propose enhancement in the MAC layer to further improvise on the results. We would also like to consider varied needs of dissimilar users for data retrieval using FCPP. We finally propose to deal with situations with mobile base stations. Conflict of interest. Authors declare that they have no conflict of interest

References 1. Dimokas, N., Katsaros, D., Tassiulas, L., Manolopoulos, Y.: High performance, low complexity cooperative caching for wireless sensor networks. Wirel. Netw. 17, 717–737 (2011). doi:10.1007/s11276-010-0311-x 2. Elbeltagi, E., Hegazy, T., Grierson, D.: Comparison among five evolutionary-based optimization algorithms. Adv. Eng. Inf. 19(1), 43–53 (2005). doi:10.1016/j.aei.2005.01.004 3. Hoyingcharoen, P., Teerapabkajorndet, W.: Fault tolerant sensor placement optimization with minimum detection probability guaranteed. In: Eighth International Workshop on the Design of Reliable Communication Networks (DRCN) (2011) 4. Huang, Yu., Cao, J., Jin, B., Tao, X., Lu, J., Feng, Y.: Flexible cache consistency maintenance over wireless adhoc networks. IEEE Trans. Parallel Distrib. Syst. 21(8), 1150– 1161 (2010) 5. Khanna, R., Liu, H., Hwa Chen, H.: Self-organization of sensor networks using genetic algorithms. In: Proceeding of International Conference on Communications, vol. 8 (2006) 6. Mollanejad, A., Khanli, L.M., Zeynali, M.: DBSR: dynamic base station repositioning using genetic algorithm in wireless sensor network. In: Second International Conference on Computer Engineering and Applications, pp. 521–525 (2010) 7. Srivastava, J.R., Sudarshan, T.S.B.: Energy-efficient cache node placement using genetic algorithm in wireless sensor networks, “Soft Computing” Journal, Springer (2014). doi:10. 1007/s00500-014-1473-8 8. Srivastava, J.R., Sudarshan, T.S.B.: A genetic fuzzy system based optimized zone based energy efficient routing protocol for mobile sensor networks (OZEEP), Appl. Soft Comput. J., Elsevier, vol. 37, pp. 863–886, December, 2015. doi:10.1016/j.asoc.2015.09.025 9. Sun, P., Ma, J., Ni, K.: A genetic simulated annealing hybrid algorithm for relay nodes deployment optimization in industrial wireless sensor networks. In: IEEE International Conference on Computational Intelligence for Measurement Systems and Applications (CIMSA) (2012) 10. Wang, Y., Li, X., Tian, J.: Blackboard mechanism based genetic algorithm for dynamic deployment of mobile sensor networks. In: International Conference on Electronic & Mechanical Engineering and Information Technology (2011) 11. Yang, L.: Determining sink node locations in wireless sensor network. In: IEEE International Conference on Systems, Man, and Cybernetics, vol. 4, Taiwan (2006) 12. Yong, M., Huludao, L., Yu, Y., Yan, W.: Optimization design of coal mine body sensor network based on Genetic Algorithm. In: International Conference on Network Security, Wireless Communications and Trusted Computing (2009). doi:10.1109/NSWCTC.2009.347 13. Youssef, W., Younis, M.: Intelligent gateways placement for reduced data latency in wireless sensor networks. In: Proceedings of the 32nd IEEE International Conference on Communications (ICC 2007). Glasgow, Scotland, UK (2007) 14. Cao, J., Zhang, Y., Xie, L., Cao, G.: Data consistency for cooperative caching in mobile environments. Computer 40(4), 60–67 (2007)

A Fast JPEG2000 Based Crypto-Compression Algorithm: Application to the Security for Transmission of Medical Images Med Karim Abdmouleh(B) , Hedi Amri, Ali Khalfallah, and Med Salim Bouhlel Research Unit: Sciences and Technologies of Image and Telecommunications, Higher Institute of Biotechnology, University of Sfax, Sfax, Tunisia [email protected], [email protected], [email protected], [email protected]

Abstract. Over the past years, the use of telecommunications and information technologies in medicine is evolving. This involves the development of the applications bound to the telemedicine and based on a network medical image transmission. Therefore, the optimization of medical application performances remains a necessity. In this paper, we propose a novel and efficient crypto-compression algorithm. This novel scheme concerning the application of a partial encryption to the JPEG2000 file format. Our algorithm is rapid, efficient, secure and it perfectly preserves the performances of the JPEG2000 compression algorithm. In addition, the proposed transmission scheme is adapted to the Telediagnostic sector and can be easily integrated in JPEG2000 coder. Keywords: Crypto-compression · JPEG2000 · RSA · Medical images Transmission · Telemedicine

1

·

Introduction

The extraordinary evolution of information technologies in the medical sector is motivated by political, professional and industrial wills. Currently, Telediagnostic is considered as the most potential sector in telemedicine [11–15]. It allows two or several medical teams to exchange medical information and to comment on it in a context of diagnosis help. This sector is the result of the creation of “proficiency poles” that is involved by the medical hyper specialization. Given the importance of this sector in the improvement of the care quality, the reduction of treatment costs and the universalization of the medical practices and knowledge, it is necessary to optimise this class of applications. To do this, both encryption and compression must be performed [4]. Various encryption schemes for JPEG2000 images have been proposed [16,17,19,26,30,35,36]. Based on the order to mix compression and encryption c Springer International Publishing AG 2018  V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8 14

A Fast JPEG2000 Based Crypto-Compression Algorithm

165

together to reduce the overall processing time [6,7,20,25,28,34], but they are either insecure or too computationally intensive and may induce degradation of coding efficiency. In fact, network communication, especially on the World Wide Web, can be intercepted. Therefore, there is a risk of placing at the disposal of eavesdroppers the secrecies of patients’ state during the transmission of medical data through a non-protected network. Thus, the encryption of the medical image is imposed to assure the deontology and the safeguard of patient medical secrecy. In particular, the asymmetric encryption can also be used to assure the authentication of medical image transmission. In addition, because of the important size of medical images after the digitalisation, we must compress them in order, first, to improve the capacity of storage, and second to optimize theirs transmission through networks (reduction in the transmission time and in the obstruction of networks). To satisfy these conditions, the classic approach consists in applying a compression algorithm to the medical image [4,27]. The output is then encrypted with an independent cryptographic algorithm. Unfortunately, the processing time in classical cryptographic algorithms is too long to satisfy the Telediagnostic required speed because this type of application is mainly used in a case of emergency. The rest of this paper is organized as follows. In the Sect. 2, the encryption algorithm RSA and the JPEG2000 still image compression algorithm are briefly described to facilitate the development of our proposed algorithm. Section 3 presents the principle of our approach. Then, the proposed encryption and decryption procedure is detailed. After that, the robustness of our algorithm against the different types of attacks is discussed. Finally, the advantages of our algorithm are enumerated. Section 4 concludes the paper.

2 2.1

Background The Asymmetric Encryption Algorithm: RSA

The RSA is a public-key cryptosystem that was invented by Rivest et al. [31]. It may be used to provide both confidentiality and authentication, and its security is based on the intractability of the integer factorization. The cryptosystem RSA utilizes two distinct keys: the public key and the private key. The first type can be communicated freely, through an insecure channel, to all the correspondents susceptible to exchange data with the possessor of the private key. On the other hand, the private key must remain confidential [2] (Fig. 1). 2.1.1 Key Generation For the creation of an RSA public key and a corresponding private key, each person X should, first, generate two large random (and distinct) primes p and q, that have roughly the same size and then compute [29]: n = p × q and φ = (p − 1)(q − 1)

166

M.K. Abdmouleh et al. Public key (n, e) Original Image

Private Key (d) Encrypted Image

Encryption

Decryption

Reconstructed Image

Insecure Channel

Fig. 1. Principle of asymmetric encryption.

Afterwards, he should select a random integer e: 1 < e < φ, such that gcd(e, φ) = 1 Next, the extended Euclidean algorithm is used to compute the unique integer d : 1 < d < φ, such that e · d ≡ 1(modφ) As result, X’s public key is (n, e) and X’s private key is d [29]. 2.1.2 Algorithm Let’s suppose that a person Y wants to send encrypted data to X who must be able to decrypt these encrypted data. If the size of n in bits is t, the encryption processing requires, firstly, the break of data into k bits of blocks and the conversion of these bocks into their decimal representation (m). To encrypt each integer m, which is in the interval [0, n−1], Y should obtain X’s public key (n, e) and compute the corresponding c: c = me mod n Convert this c into binary representation, redo this operation with all the k bits of blocs, and finally send the encrypted data to X. After doing all the necessary conversions and breaks, X should use the private key d to recover m from c [29]: m = cd mod n The RSA adds a very significant benefit: it can serve to authenticate the sender of the data (e.g. a digital signature) [9,10,21]. In fact, the RSA keys are complementary: if one encrypt with a public key, it must use the private key for recovering the encrypted data (Fig. 2). Inversely, if one encrypt with a private Private Key (d) Original Image

Public key (n, e) Authentic Image

Encryption

Decryption

Reconstructed Image

Insecure Channel

Fig. 2. Principle of authentication with RSA algorithm.

A Fast JPEG2000 Based Crypto-Compression Algorithm

167

key, it must use the public key to recover the encrypted data. Therefore, when X encrypts his own image with his own secret key, Y will be able to use the X’s public key to decrypt the image, and in this case, Y is certain that it is X who sent him the image because he is the only person who possess the secret key. 2.1.3 Security The security of the RSA algorithm is generally equivalent to the hardness of the factoring problem. In fact, to find the private key d, the only possibility is the factorization of n [33]. However, we need to be careful in the choice of the key size for RSA because of the increase in computing power and the continuing refinements of the factoring algorithms [33]. The last factoring record was achieved on December 12, 2009 when a sixinstitution research team led by T. Kleinjung threw down a challenge to decrypt RSA-768 number. The effort took almost 2000 2.2 GHz-Opteron-CPU years according to the submitters, just short of three years of calendar time [22]. 2.2

JPEG2000

The JPEG2000 standard was created by the Joint Photographic Experts Group (JPEG) [1], also denominated as ISO/IEC 15444. It’s an image compression algorithm that allows great flexibility, not only for the compression of images, but also for the access to the compressed data, such as the several progressive decoding modes: by resolution, by quality, by position or by region. Currently, this standard is divided into fourteen distinct parts, where Part 1 [18] describes the core coding system and Part 2 its extensions. However, only the first part reached the International Standard (IS). This part describes mainly the decoding algorithm and the compressed data format: the codestream is the result of a juxtaposition of headers containing coding parameters and bit streams (a compressed form of image data). Since our approach is about the application of a partial encryption in the JPEG2000 file format, only a short preview about the codestream building will be presented in the next paragraph. The JPEG2000 file format (JP2 format): This format provides a foundation for the storing application of the specific data (metadata). A JP2 File represents a collection of boxes (Fig. 3) which design a building block defined by a unique box type and length [1]. As shown in Fig. 4, four fields compose a box: the LBox, TBox, XLBox, and DBox fields. The LBox field specifies the length of the box in bytes. The TBox field indicates the type of box (i.e., the nature of the information contained in the box). The XLBox field is an extended length indicator that provides a mechanism for specifying the length of a box the size of which is too large to be encoded in the length field alone. If the LBox field is 1, then the XLBox field is present and contains the true length of the box. Otherwise, the XLBox field is not present [1]. The DBox field contains data specific to the particular box type. Some of the JP2 File boxes are independent, and some of those boxes contain other boxes.

168

M.K. Abdmouleh et al.

JP2 File JP2 Signature box Profile box JP2 header box (superbox) Image header box Bit Per Component box Component Definition box Color Specification box 0 Color Specification box n-1 Palette box Resolution box (superbox) Capture resolution box Definition display resolution box Contiguous codestream box 0 Contiguous codestream box m-1 IPR box XML boxes UUID boxes UUID Info boxes (superbox) UUID List box Data Entry URL box

Fig. 3. Conceptual structure of JP2 File.

LBox TBox

XLBox (if required)

DBox

32 bits 32 bits

64 bits

Variable

Fig. 4. Box structure.

Box 0

Lbox 0

Box 1

Lbox 2

Lbox 3

Box 2

Box 3

Lbox 1

Box 4

Lbox 4

Fig. 5. Illustration of box length.

As a matter of terminology, a box that contains other boxes in its DBox field is referred to as a superbox. The binary structure of a file is a contiguous sequence of boxes (Fig. 5).

A Fast JPEG2000 Based Crypto-Compression Algorithm

3

169

The Fast Crypto-Compression

In this section, we will develop the proposed JPEG2000 based cryptocompression algorithm. This novel scheme concerning the application of a partial encryption to the JPEG2000 file format (Fig. 6). Firstly, the principle of the novel approach, on which our algorithm is based, is presented. Second, the proposed robustness of our algorithm to the different type of attacks is proved. Finally, the advantages of our crypto-compression technique are enumerated.

Original Medical Image

Reconstructed Medical Image

JPEG2000 Compression

JPEG2000 Decompression

Partial Encryption in JPEG2000 File Format

Partial Decryption in JPEG2000 File Format

Insecure Channel

Fig. 6. Scheme of our approach.

3.1

Principle of Our Approach

If we examine the JP2 File structure attentively, we conclude that most of the information in this type of file is independent (Structure in packet). Therefore, to design an efficient crypto-compression algorithm, we must touch the totality of the file structure. Since the JP2 File format is composed of boxes, and the first 32 bits (Lbox) of every box contain the box length information. Actually, the basic idea of our approach consists of encrypting only the Lbox part of every box in the JP2 File format excepting the boxes included in a superbox. Consequently, it will be impossible to recover the original data because the correct box length is unknown. As a result, the image is protected. 3.2

Encryption Algorithm

In Sect. 2, it is noted that the principle of the RSA algorithm consists of encrypting a blocs of K bits (where K is the length in bits of the RSA key). Hence, if one uses the RSA algorithm for encrypting, the length L of the data that can be encrypted, in bits must be a multiple of K : L = n × K (n is an integer)

170

M.K. Abdmouleh et al.

Since it is possible that the Lbox bit sequence is not a multiple of K bits, we propose the following encryption procedure: • Regrouping all the Lboxes in a structure of a contiguous bit sequence without changing their order (Fig. 7a). This (32 × B) bit sequence must be placed at the beginning of the JP2 File (where B represents the number of JP2 File boxes excepting those that are included in a superbox). • Adding 32 bits in order to indicate the length of the Lbox sequence length (32 × B bits) because the recovering of the original data requires the transmission of the Lbox sequence length. This few added bits must also be placed at the beginning of the JP2 File but before the Lbox sequence (Fig. 7b). • Determine an integer N that verifies: N < K and (32 + 32 × B + N ) is a multiple of K Then, we encrypt the (32+32×B+N ) bits with the RSA encryption algorithm (Fig. 7c). For decryption the encrypted data in the reception, we must: • Decrypt the K bits from the beginning of JP2 File. The first 32 decrypted bits allow the determination of the Lbox sequence length (32 × B bits) (Fig. 7d). • Search the integer N, with the same way that is employed at the encryption stage. Once N is determined, we must decrypt the remaining encrypted data: [(32 + 32 × B + N ) − K] bits. • Delete the first 32 bits that indicate the Lbox sequence length. • Place the new first 32 bits (corresponding to the Lboxes of the first Box in JP2 File) just after the Lbox bit sequence. Consequently, the length of the first box is known. The second 32 bits are placed just after the end of the first box. Since the length of the second box is known, we can place the next 32 bits at the end of the second box and so forth until all the Lboxes in their original places (Fig. 7e). After the decryption stage, the original image is recuperated by decoding the resulting JP2 File with any classical JPEG2000 decoder. 3.3

Security

In this part, we discuss the robustness of our crypto-compression algorithm against different attacks through the cryptanalysis that is the study of breaking encryption algorithms. In this discussion, we assume that the cryptanalyst have full access to the description of the algorithms, as well as full access to the insecure channel through which the image is transmitted [33]. The security of our image cryptosystem will be analysed for the following four types of attacks, such as exhaustive key search attack, cipher image-only attack, known-plain image attack, and chosen-plain image attack. In the Exhaustive key search, the cryptanalyst tests each of the possible keys one at a time until the correct plain image is recognized [8]. Since we use

A Fast JPEG2000 Based Crypto-Compression Algorithm

171

LBoxes (32 bits)

Box 1

(a)

A supplementary 32 bits (The Lboxes bit sequence length)

(b) N bits

x.K bits (x is an integer)

Encrypted Data

(c)

K bits

(d)

(e)

Fig. 7. The proposed encryption and decryption stages.

the RSA keys for the encryption and decryption of the box bit sequence in our algorithm, the break of our algorithm with this type of attack is equivalent to the test of all the possible RSA Keys. In the cipher image-only attack, the cryptanalyst has access only to several images that are encrypted with the same key. The cryptanalyst attempts to recover the corresponding plain image or the decryption key [8].

172

M.K. Abdmouleh et al.

Since it is impossible to recovering the RSA decryption key, and it is impossible to recovering any information about the original data without the acquaintance of the length of the JP2 File boxes (Lbox), our algorithm is robust against this type of attacks. In the Known-plain image attack, the cryptanalyst has access to the cipher image and the corresponding plain image of several images encrypted with the same key. The cryptanalyst attempts to recover the key or to design an algorithm to decrypt any image that is encrypted with the same key [33]. In the case of our algorithm, the encryption is made on a bloc of K bits (RSA principle). The result of the RSA encryption of two original K bits blocs that contain few different bits is two K bits blocs that are completely different. Since the original bit sequence of two similar images contain certainly few different bits, it is impossible to find a relation between the original and the encrypted blocs. Consequently, it is useless to cryptanalyse our algorithm with a know-plain image attack. In the Chosen-plain image attack, the cryptanalyst is allowed to choose the plain image that is encrypted, and observe the corresponding cipher image. The cryptanalyst’s goal is the same as that in a known-plain image attack. The cryptanalyse of our algorithm is equivalent to that of the bit sequence that is encrypted with RSA encryption algorithm. As a consequence, the break of our crypto-compression algorithm with the chosen plain image attack is equivalent to that of the RSA technique with this type of attack. 3.4

Advantages of Our Scheme

• Our algorithm is rapid. In fact, we encrypt only the Lbox part (32 bits) from the whole information that can be contained in the JP2 File. Therefore, our approach allows an enormous reduction of the encrypting-decrypting processing time. • Our algorithm perfectly preserves the performances of the JPEG2000 compression algorithm [32] because the encryption is made after all the compression stages, and therefore, it does not tamper with the compression ratio. The 32 bits, which are added to communicate the Lbox bit sequence length, are too negligible, therefore, one cannot consider this fact as an inconvenient. • Moreover, our algorithm is secure. In fact, it is demonstrated in Sect. 3 that the security of our crypto-compression system is equivalent to that of the RSA algorithm. This security depends essentially on the key length. According to the RSA factoring records, if we use a key size in the range of 1024 to 2048 bits, we can affirm that our algorithm is robust to the different attacks for many years. • Our algorithm is adapted to the medical image transmission in Telediagnostic [3,5,23,24]. In addition, since RSA permits the authentication of the sender of the medical image [9,10,21], we can assure a secure and an authentic medical image communication with our algorithm.

A Fast JPEG2000 Based Crypto-Compression Algorithm

4

173

Conclusion

In this paper, we proposed a new crypto-compression algorithm that significantly reduces the encryption and decryption time and can assure a secure and an authentic transmission in medical image communication. Based on the structure of the JP2 File, which is based on a succession of box and superbox, we developed, in this work, an original approach concerning the encrypting in JP2 File only the part that indicates the length of the boxes (the Lbox). Given the originality and the efficiency of this approach, the proposed algorithm, which is derived from, is rapid, secure, and preserves the excellent performance of JPEG2000 algorithm and adapted to the medical image transmission in Telediagnostic sector. In addition, our algorithm can be applied in many other telecommunication sectors like Tele-expertise, secure broadcasting, and distributed image retrieval.

References 1. ISO/IEC 1.29.15444-1, “JPEG2000 part i final committee version 1.0”, September 2004 2. PKCS#1 v2.2, RSA cryptography standard, RSA Laboratories (2012) 3. Amri, H., Hanna, F., Lapayre, J.C., Khalfallah, A., Bouhlel, M.S.: Repro: a new reduction/expansion protocol to increase the performance of image transmission in medical telediagnosis platforms. Biomed. Eng.: Appl. Basis Commun. 27(06), 1550054 (2015) 4. Bouhlel, M.S., Kammoun, F., Garcia, E.: An efficient DCT-based cryptocompression scheme for a secure and authentic medical image transmission. J. Test. Eval. Appl. Sci. Eng. 34(6), 459–463 (2006) 5. Chaabouni, I., Fourati, W., Bouhlel, M.S.: Using ROI with ISOM compression to medical image. IJCVR 6(1/2), 65–76 (2016) 6. Chang, C.C., Hwang, M.S., Chen, T.S.: A new encryption algorithm for image cryptosystems. J. Syst. Softw. 58(2), 83–91 (2001) 7. Chang, H.K.C., Liu, J.L.: A linear quadtree compression scheme for image encryption. Signal Process.: Image Commun. 10(4), 279–290 (1997) 8. Cheng, H., Li, X.: Partial encryption of compressed images and videos. IEEE Trans. Signal Process. 48(8), 2439–2451 (2000) 9. Cramer, R., Damgard, I.: New generation of secure and practical RSA-based signatures. In: 16th Annual International Cryptology Conference, pp. 173–185. Springer Berlin Heidelberg, Berlin, Heidelberg (1996) 10. Cramer, R., Shoup, V.: Signature schemes based on the strong RSA assumption. ACM Trans. Inf. Syst. Secur. 3(3), 161–185 (2000) 11. Dey, N., Bose, S., Das, A., Chaudhuri, S.S., Saba, L., Shafique, S., Nicolaides, A., Suri, J.S.: Effect of watermarking on diagnostic preservation of atherosclerotic ultrasound video in stroke telemedicine. J. Med. Syst. 40(4), 1–14 (2016) 12. Dey, N., Mukhopadhyay, S., Das, A., Chaudhuri, S.S.: Analysis of P-QRS-T components modified by blind watermarking technique within the electrocardiogram signal for authentication in wireless telecardiology using DWT. Int. J. Image, Graphics Signal Process. (IJIGSP) 4(7), 33–46 (2012)

174

M.K. Abdmouleh et al.

13. Dey, N., Nandi, B., Das, P., Das, A., Chaudhuri, S.S.: Retention of electrocardiogram features insignificantly devalorized as an effect of watermarking for a multimodal biometric authentication system. In: Advances in Biometrics for Secure Human Authentication and Recognition, pp. 175–212 (2013) 14. Dey, N., Pal, M., Das, A.: A session based blind watermarking technique within the NROI of retinal fundus images for authentication using DWT, spread spectrum and harris corner detection. Int. J. Mod. Eng. Res. (IJMER) 2(3), 749–757 (2012) 15. Dey, N., Samanta, S., Yang, X.S., Das, A., Chaudhuri, S.S.: Optimisation of scaling factors in electrocardiogram signal watermarking using cuckoo search. Int. J. BioInspired Comput. 5(5), 315–326 (2013) 16. Fang, J., Sun, J.: Compliant encryption scheme for JPEG2000 image code streams. J. Electron Imaging 15(4), 043,013–043,013–4 (2006) 17. Grangetto, M., Grosso, A., Magli, E.: Selective encryption of JPEG2000 images by means of randomized arithmetic coding. In: IEEE 6th Workshop on Multimedia Signal Processing, pp. 347–350 (2004) 18. Grosbois, R., Santa-Cruz, D., Ebrahimi, T.: New approach to JPEG2000 compliant region of interest coding. In: SPIE’s 45th Annual Meeting, Applications of Digital Image Processing XXIV, vol. 4472, pp. 267–275 (2001) 19. Gu, G., Ling, J., Xie, G., Li, Z.: A chaotic-cipher-based packet body encryption algorithm for JPEG2000 images. Signal Process.: Image Commun. 40, 52–64 (2016) 20. Jones, D.: Application of splay trees to data compression. Commun. ACM 31(8), 996–1007 (1988) 21. Katz, J.: Digital Signatures. Springer, US (2014) 22. Kleinjung, T., Aoki, K., Franke, J., Lenstra, A.K., Thom´e, E., Bos, J.W., Gaudry, P., Kruppa, A., Montgomery, P.L., Osvik, D.A., Te Riele, H., Timofeev, A., Zimmermann, P.: Factorization of a 768-bit RSA modulus. In: Proceedings of the 30th Annual Conference on Advances in Cryptology, pp. 333–350 (2010) 23. Lazrag, H., Naceur, M.S.: Despeckling of intravascular ultrasound images using curvelet transform. In: 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications (SETIT), pp. 365–369 (2012) 24. Lazrag, H., Naceur, M.S.: Wavelet filters analysis for speckle reduction in intravascular ultrasound images. In: 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications (SETIT), pp. 375–379 (2012) 25. Li, X., Knipe, J., Cheng, H.: Image compression and encryption using tree structures. Pattern Recognit. Lett. 18(11–13), 1253–1259 (1997) 26. Lian, S., Sun, J., Zhang, D., Wang, Z.: A selective image encryption scheme based on JPEG2000 codec. In: 5th Pacific-Rim Conference on Multimedia PCM, pp. 65–72 (2004) 27. Lima, J., Madeiro, F., Sales, F.: Encryption of medical images based on the cosine number transform. Signal Process.: Image Commun. 35, 1–8 (2015) 28. Matias, Y., Shamir, A.: A Video Scrambling Technique Based On Space Filling Curves. In: Proceedings of Advances in Cryptology (CRYPTO), pp. 398–417. Springer Berlin Heidelberg, Berlin, Heidelberg (1988) 29. Menezes, A.J., Van Oorschot, P.C., Vanstone, S.A.: Handbook of Applied Cryptography. Taylor & Francis (1997) 30. Modrzyk, D., Staworko, M.: A high-performance architecture of JPEG2000 encoder. In: Proceedings of the 19th European Signal Processing Conference (EUSIPCO), pp. 569–573 (2011) 31. Rivest, R.L., Shamir, A., Adleman, L.: A method for obtaining digital signatures and public-key cryptosystems. Commun. ACM 21(2), 120–126 (1978)

A Fast JPEG2000 Based Crypto-Compression Algorithm

175

32. Saidani, T., Atri, M., Said, Y., Tourki, R.: Real time FPGA acceleration for discrete wavelet transform of the 5/3 filter for JPEG 2000. In: 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications (SETIT), pp. 393–399 (2012) 33. Schneier, B.: Applied Cryptography: Protocols, Algorithms, and Source Code in C. Wiley, New York (1996) 34. Wang, C., Ni, J., Huang, Q.: A new encryption-then-compression algorithm using the rate-distortion optimization. Signal Process.: Image Commun. 39, 141–150 (2015) 35. Wu, H., Ma, D.: Efficient and secure encryption schemes for JPEG2000. In: IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), vol. 5, pp. 869–872 (2004) 36. Yan, S., Lin, Q.: Partial encryption of JPEG2000 images based on EBCOT. In: International Conference on Intelligent Control and Information Processing (ICICIP), pp. 472–476 (2010)

IoThings: A Platform for Building up the Internet of Things Andrei Gal(&), Ioan Filip, and Florin Dragan Department of Automation and Applied Informatics, University Politehnica of Timisoara, Bvd. V. Parvan, no. 2, 300223 Timisoara, Romania [email protected], {ioan.filip,florin.dragan}@aut.upt.ro

Abstract. Given the multitude and diversity of smart devices surrounding us, along with the variety of applications distributed on these smart devices, this paper proposes the architecture for a platform that will be used to interconnect different smart devices in terms of hardware capabilities and software features, making feasible the concept of Internet of Things (IoT). The paper defines the architecture of the platform, expanding the functionality and correlation between modules that make up the platform, and proposes a software representation of a Thing and also a software representation of a Context as used in the IoT network. Using both software representations, platform modules manage tasks that discover all available devices in a surrounding area and create proper context for devices to collaborate, schedule smart device communication and establish communication between available devices in the network. The great advantage of the proposed solution is that it provides an open platform, extensible from an architectural point of view with the possibility to integrate with new smart devices in order to enlarge the IoT network. Keywords: Internet of things Software architecture




Smart devices




Context-aware




Platform




1 Introduction A mobile device can be described as being a computational device having a hardware structure similar to that of a personal computer (PC) and also a processing power compared to it. The main design feature of all mobile devices is that they are that small that a person can easily move a device from one place to another. Usually their dimensions are basically compared to that of a human hand. The evolution of mobile devices to what we now define as being a mobile device took place in quite a small period of time. The impact of the technological development in this direction was great on society affecting it on different levels such as economics, social media, commerce, press, culture, and nevertheless having the greatest impact on the communication level in the society. Below there are defined 5 periods in the evolution of mobile devices:

© Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_15

IoThings: A Platform for Building up the Internet of Things

177

– The Brick (1973–1988) - this is the first known era in the mobile telephony as in this period the fundamentals of future cell phone communication were being put in place along with the appearance on the market of the first mobile phone (cell phone) invented by Martin Cooper of Motorola company in 1983 [1]. – The Candy Bar (1988–1998) - this era represents one of the biggest leaps in the mobile technology. During this period the design of the cell phones changed drastically both because the usage of second-generation (2G) technology and the increased density of cellular sites [1]. – The Feature Phone (1998–2008) - Taking in consideration the technological advancement to the mobile devices, this era is considered having the lowest impact. Most of the changes occurred to the mobile devices were at a software level as software features were developed to make use of the available hardware such as music listening applications, photo taking applications, game applications. The most groundbreaking change was the introduction of the internet and the addition of GPRS modules along with the usage of the new 2.5G network [1]. – The Smartphone (2002–present) - This era overlapped the one before as there is no clear difference between devices developed during the previous era and this one. Though introducing the internet on a mobile phone might be the event that triggered the so called smartphone era, the real route was made when introducing a common operating system (OS) on each of the devices developed. This addition led to the possibility of developing OS based application which gave birth to the concept of smartphone (a phone that is as smart as a PC). Alongside the development on a large scale of smartphones companies started to extend the area of development creating other smart devices such as smart TV’s, smart watches, smart cards, smart boards, smart pads, smart fridges, smart locks etc. [1]. – The Touch (2009–present) - This era overlaps the one before and is related to the introduction of the touchscreen technology on smart mobile devices but also on simple smart devices. The first touchscreen technology was implemented by Apple Inc. on their first iPhone and since then every smartphone uses it. As with smartphone technology, touchscreen technology began to be used increasingly more and applied to a number of existing smart devices, nowadays having touchscreen smart TV’s, touchscreen smart watches, touchscreen fridges touchscreen laptops etc. [1].

2 An Overview of the Used Concepts 2.1

Internet of Things (IoT)

During the mobile device development evolution most of the device could communicate between them only by using the GSM network, therefore the devices were mainly used as mobile phones with no possible way to interconnect them. This was mostly related to the fact that there was a limitation of the communication environments that the mobile devices possessed.

178

A. Gal et al.

Along with the development of smartphones with Wi-Fi and Bluetooth modules applications that use these modules to exchange data were more often developed. Most applications exchanged address book information such as contact entries, text messages, and call history but there were also applications developed to exchange files such as music or photo files. With the development of these devices and applications the need of having a live interconnection for exchanging relevant data between the existing devices has increased to a point that the concept of Internet of Things appeared. The Internet of Things can be defined as being a network on a global scale for interconnecting smart devices (Things) around the world. Making a comparison with the internet as we know it where people are the agents of communication and data transferred between people is mostly referring general social information, the IoT uses Things as agents of communication, while the data transferred between the Things represent relevant data for the Things engaged in the process of communication. As the concept is being still in early stages there are several possible interpretations of the Things that make up the IoT. Considering the current development of smart devices focused mostly on the home automation services, a possible basic IoT infrastructure can be seen in Fig. 1. All smart devices in Fig. 1 play the role of Things in the IoT infrastructure [2, 3].

Fig. 1. Basic IoT infrastructure

In particular, the simplest use of the IoT, by referring to Fig. 1 shown above, could be that of a user being able to just power on/off the smart TV from a simple TV Manager application the user has installed on the smartphone he owns. When referring to the term home automation a more complex infrastructure of the IoT could be represented by involving a lot more Things in the process of automating the home. Things like temperature sensors, pressure sensors, humidity sensors or light sensors embedded in equivalent smart devices can give feedback upon action that could be taken to adjust the ambient comfort in the home. The actions taken to adjust the environment could be done by Things like windows’, air conditioners’, dehumidifiers’, electric heaters’, louvers’ all developed as smart devices that could be controlled remotely like in Fig. 2 [3].

IoThings: A Platform for Building up the Internet of Things

179

Fig. 2. Complex IoT infrastructure

By taking in consideration various points of interest, the IoT infrastructure can be extended by adding Things relevant in fields of study like health care, smart buildings, waste management, smart roads, social media and a lot more fields where it can be possible to build a level of abstraction so that Things can be developed to serve purposes like people’s security, replacement of human actions, increased comfort level in an environment, reducing the possibility of human error, etc. In an ideal IoT world, smart devices will serve users by doing autonomous work behind the scenes, accomplishing the ultimate purpose for which such an enlarged IoT infrastructure is created, that being interconnectivity between devices and services that work together to do something useful for mankind [3, 4]. 2.2

Context-Aware Computing

To present in a clear light the essential aspects of the proposed platform it can be mentioned that like many current architectures proposed for implementation of IoT infrastructure, IoThings was also designed taking in consideration the paradigms of context-aware computing, especially emphasizing on the main attribute of any context-aware system namely the adaptation of the system to its location, to the people and objects in the system or connected to the system, and also to the changes of the objects over time [5, 6]. The concept was developed based on the ubiquitous computing paradigm proposed in the early 1990’s and it refers to a suite of mobile systems that can take information from their environment and adapt to it by taking appropriate decisions. The core

180

A. Gal et al.

elements of these context-aware systems are represented by the context information and the actions (decisions) taken to adapt the mobile systems to the context at a certain point in time and space. The platform presented in this paper follow the lines drawn by both the context-aware computing concept and IoT concept, proposing a modular platform architecture that encapsulated core aspects from both concepts presented above with the purpose to create a context-aware IoT network where devices can interconnect forming the Internet of Things [7, 8].

3 IoThings - Software Level Concepts IoThings is a platform designed to support a series of smart devices that are different in terms of hardware functions and could run under different or same operating systems. The platform is designed, to have independent modules overlapping each of the functions of the platform, altogether coordinating the devices to adapt to the environment they reside in, for creating the proposed context. Due to the focus on the context-aware computing concept, the platform is structured around the term of context information. As the platform is being designed for implementing an IoT solution platform, all the smart devices that use the platform are represented as entities belonging to a certain context, and all relevant information about a smart device is translated in a context entity information [9, 10]. 3.1

Software Representation of a Thing

As defined by the IoT concept, a Thing is represented in the real world by a smart device and its main characteristic is that it can interconnect with other Things for exchanging data, this way creating a large network of Things which eventually can create the Internet of Things. Having a multitude of various Things that are being different in terms of hardware structure, the platform defines a set of hardware functionalities that a Thing can have. For example, Things can get measures from the environment, can record sounds/videos, can move, can carry other Things, can sense proximity, can start/stop engines etc. A Thing being represented as a class at a software level, each of the hardware functionalities will be defined as an interface, so for the abstraction of each function a Thing has, it will have to implement that particular interface. In Fig. 3 is presented a class diagram for the interfaces the platform supports. It should be mentioned here that the class diagram can be extended to include any other possible hardware features a Thing can own but also adapt existing features to requirements of an accepted Thing for the platform [11]. As seen in Fig. 3, a Thing represented at a software level in the IoThings platform contains the sum of all functions that that Thing is supporting at a hardware level. Going onwards the platform also reduces to a level of software abstraction the general information a Thing can have like: name, physical form, physical dimension, operating system installed on, operating system version, software version etc. As depicted in Fig. 4 the platform exposes a generic interface that the Thing will have to implement in order to work with the platform.

IoThings: A Platform for Building up the Internet of Things

181

Fig. 3. Defined interfaces for the software representation of a Thing’s functions

Fig. 4. Defined interfaces for the software representation of a Thing’s general information

The general information interface contains nine properties as below: – Name – property that defines the name of the Thing. – OperatingSystem (OS) – property that defines the operating system under which the software embedded in the Thing runs. This property is relevant in terms knowing how to handle the communication between Things that have different operating systems installed. – OperatingSystemVersion – property that gives the platform the information regarding the version of the OS and as like the one above it’s important in communication between Things running different OS/OS versions. – Form – property that represents the physical form of the Thing. As it can be seen in Fig. 4 this can have several values like: Circle, Sphere, Rectangular, Square, Cube, Triangle. – Height – property that is set to the real height of the Thing. – Width – property that is set to the real width of the Thing. – Depth – property that is set to the real depth of the Thing. – Weight – property that is set to the real weight of the Thing. – IsRound – this property is set if the Thing represented has a round shape. – Radius – this property comes with the one above and it represents the actual radius of the Thing in case it has a round shape.

182

A. Gal et al.

The function of a Thing that can move is represented by the IMoveFunction interface and it contains the following: – IsMoving – property that is set whenever the Thing is in motion. This is relevant in the context of having to give the Thing a command to move knowing it is already in motion. – MoveToLocation – this method is used to give the Thing the command to move to a specified location in the available space. – LocationChanged – this event is triggered by the Thing whenever its location has been changed. This way the platform is able to locate each of the Things at a specified moment in time. To represent the capacity of a Thing to take measurements from the environment it resides in the IMeasureFunction interface propose below members: – IsMeasuring – property specifying the fact that the Thing is currently taking measurement from the environment. The property is useful when using Things that work with big data values or have time consuming sensors. – MeasuredValue – property that retains the measured value from the environment at a certain point in time. – MeasureType – this property is set to the relevant type of measurement the Thing deals with. As seen in Fig. 5 the property can have several values like: – MeasureUnit – property is set to the relevant unit of measurement used by the MeasureType property, and it can take values like:

Fig. 5. Measure types and measure units used by IMeasureFunction interface

– GetMeasurement – the method is used to retrieve the last value read by the Thing from the environment. This will also update the MeasuredValue property. – NewMeasuredValue – event is used to be triggered whenever a new value has been measure from the environment. The ability of a Thing to execute commands that have an effect on the environment is represented by the ICommandFunction and it has the following composition: – CommandType – property that can be set with the following values as in Fig. 6: Start, Stop, Open, Close, Lock, Unlock, Drag, Move, Push, Pull, Cut.

IoThings: A Platform for Building up the Internet of Things

183

– Status – property that is set with the result of the command’s execution, as for to know whether the command ended successfully has entered some state relevant for the context in which the Thing is. The values possible for this property can also be seen in Fig. 6. – IsExecuting – this property is set whenever a command starts its execution and is reset when the command end the execution. – ExecutionMessage – property will retain the message thrown by the command after finishing the execution. – AsyncExecuteCommand – the method is used to start the asynchronous execution of the implemented command. As the execution of a command can be time consuming this method offers the possibility to run the execution on a separate thread in case necessary in the context. – ExecuteCommand – the use of this method implies the execution of the command awaiting for it to finish, blocking the thread from where it is called. – CommandStatusChanged – this event is fired up by the Thing when the status of the command in execution is changed.

Fig. 6. Command types and command status used by ICommandFunction interface

For the video ability of a Thing, the IVideoFunction interface has to be implemented having the structure below: – CanPlayVideo – property that defines whether the Thing can play or not videos. – CanRecordVideo – property that defines whether the Thing can record or not videos. – NumberOfCameras – property set to the number of available cameras that the Thing has to offer. – OptimalResolution – property set with the optimal resolution for the video. This will be used for the Thing to record/play video by setting the resolution to the one in the property. – GetVideoStream – this method is used to get the video stream from a certain camera and use it as output for any of the applications using the platform and using live streaming feature. – StartVideoPlayer – this method is used to tell the Thing to start playing a video in case it has the CanPlayVideo property set. – StartVideoRecording – this method is used to tell the Thing to start recording video in case it has the CanRecordVideo property set.

184

A. Gal et al.

The structure of the IAudioFunction, making available the audio representation at the software level for a Thing consists of: – HasMicrophone – property set when a Thing has a microphone incorporated for the possibility to record audio. – HasSpeakers – property set if a Thing has speakers available to play audio sounds. – SoundDevice – property specifying the sound device used by the Thing. – GetAudioStream – this method is used to get the audio stream from a certain device and use it as output for any of the applications using the platform and using live audio feature or voice recording feature. – StartAudioPlayer – this is used to start playing and audio file in case a Thing has HasSpeakers property set. – StartAudioRecording – this method is used to start recording audio is a Thing has HasMicrophone property set. Taking in consideration the context information core element of the IoThings platform, all Things known by the platform have to be included in a context so that the context-awareness paradigm is respected. To include all Things in a certain context the platform integrates the Thing concept into the context information concept by creating a context entity that encapsulates any Thing known by the platform. In order to do a better drawing of the context concept any context entity will consist in two major parts as seen below in Fig. 7.

Fig. 7. Context Entity structure

As it can be seen in Fig. 7 a Context Entity is composed of a Location Context and a Thing Context. The Thing Context is represented by the software representation of a Thing which is described above, as the Location Context represents the location of the Context Entity in the context space bordered by the IoT network that is built with the use of IoThings platform. Particularly speaking the Location Context consist of relevant information regarding the position in time and space where a specific entity is. From a structural point of view, the Location Context consists of: – GPS coordinates – if possible this is to be provided by the Thing in the Context Entity, or could be provided by entities from the same context. – Context reference – a reference to the context that the entity belongs to. – Time reference – a time reference based on where the context resides in terms of geo-location. It can contain data as local time, local day, local month [12].

IoThings: A Platform for Building up the Internet of Things

3.2

185

Software Representation of a Context

To be able to have an idea of how IoThings platform uses the Context concept to create an abstraction at the software level we’ll highlight some definitions of the concept. A Context could be seen as environmental information that is part of an application’s operating environment and that can be sensed by the application. This typically includes the location, identity, activity and state of people, groups and objects that are part of the environment. Another possible definition of a Context could be that of the interaction between humans and computers in socio-technical systems that takes place referring to the physical and social situation in which computational devices and environments are embedded. The context is determined by the people involved (including their background knowledge and their intentions), the objective of the interaction (including the tasks to be carried out), and the time and place where the interactions occur [13]. Given the two definitions above, IoThings platform defines the Context as being a self-aware environment, where entities different in terms of functional abilities can interact between them or with people by giving feedback with a certain status of the environment at a point in time, or accepting commands if necessary, to adjust the environment for serving a predefined scope. Having this in consideration the software representation of a Context from the IoThings platform point of view consists in elements like in Fig. 8: – Name – property set with the name of the Context. – Location – a property that defines the geographical coordinates where the Context is settled. – Scope – property set with the predefined scope of the Context. – Status – property giving the current status of the Context. Possible status of the Context can be determined by a full analysis of the Scope of the Context. – NetworkID – property set with the ID of a certain network (if any) where the Context establishes communication between entities owned. – AvailableContextEntities – a list of Context Entity objects that are residing in the specified Context. These entities make up the Context itself by adapting and adjusting Context variables to achieve the predefined Scope. – SelfAwarePercent – property which set the percentage of self-awareness of the Context. The percentage is to be calculated depending on several factors like number of autonomous entities in the Context, number of human interactions in achieving the Scope, number of errors in executing a command in the Context etc.

Fig. 8. Software representation of a Context

186

A. Gal et al.

4 IoThings Architecture IoThings platform being designed for working with Things in terms of Context Entities, and respecting the used concepts, its architecture is composed of modules that provide functionalities like communication with any known Context Entity, defining a Context using the available entities, managing contexts and context entities, all modules making use of the software representations defined by the platform in order to interfere with the smart devices inhabiting a real context environment. The platform’s architecture, represented at the modules level is depicted in Fig. 9.

Fig. 9. IoThings architecture

a. Context Entity Service - this module represents a service exposed by the platform that gives the possibility for any Context Entity to communicate with the platform. Through this module a smart device that is defined as a possible Context Entity will be known by the platform as an available Context Entity and can be used in a proper Context based on the location and the hardware functions that the device is capable of. Also the service can be used for communication between different devices that cannot communicate due to hardware/software interoperability problems. b. Context Entity Analyzer – this module deals with the analysis of all entities in terms of both location context and functions that the entity is capable of, resulting in the creation of a new Location Context where the analysed entity will be included or in the attachment of the analysed entity to an already existing Location Context. c. Context Entity Manager – this module represents the management level of the platform related to the entities communicating with the platform, and it performs three major functions: extracts all possible entities that can communicate with the platform into a Contact Entity Container, establishes communication between the external applications available to interact with a certain context via IoThings and relates any defined Context with an external application or with any of the entities owned by the Context. d. Context Manager – this module represents the management level of the platform related to the contexts that are defined by it. The module communicates with the Context Entity Analyzer and based on the result of the analysis manages the

IoThings: A Platform for Building up the Internet of Things

187

resulting Location Context in order to create a new Context or manage existing ones in a Context Container. Also this module provides information for any of the external applications regarding any context created and managed by the platform. e. Application Service – this module acts as a service exposed by the platform in order to allow external applications to communicate with it. The applications are used by people in certain environments to either get feedback from the Things in the environment, or to command Things that can be triggered actions on.

5 Conclusions This paper presents a platform which is able to manage smart devices that have different hardware functions and software installed components, by putting together two abstractions at a software level for concepts defined in the Internet of Things and Context-Aware Computing fields of interests. A Thing is represented by the platform as a class that implements an interface related to general information that a Thing has, and also interfaces related to different functions that a Thing possesses. A Context is represented by the platform as a class that puts altogether relevant information that defines a context-aware system like: location, scope, status, context entities belonging to the context and self-awareness. The platform architecture is designed to communicate with all Things that are defined respecting the proposed software representation of a Thing, in order to create a container with all available Things for the platform. The platform will extract Things from the container, and by subjecting them to an analysis based on the location of the Thing and its capabilities, will create Contexts that will serve a predefined scope. The platform will manage both Things and Contexts in order to create an interconnection between the two terms by attaching to a Context as many Things as needed for serving the scope for which the Context was defined. The connection will be translated in real situations, where smart devices forming a real environment, will adapt to be able to maintain/achieve a certain desired state of the environment. Also the platform will tend to increase the degree the self-awareness for each of the Contexts defined by emphasizing the use of the Things in an autonomous mode. The platform makes available an application level where external applications can communicate with the platform for getting feedback from a certain Context or for sending commands to Things that act as actuators in the Context. In particular, the application level enables the human factor the possibility to intervene by just visualizing the state of the Context or by acting on some Things in that certain Context. Acknowledgments. This work was developed through the Partnerships in Priority Areas PN II, with the support of ANCS, UEFISCDI, Project No. 36/2012, Code: PN-II-PT-PCCA2011-3.2-1519.

188

A. Gal et al.

References 1. Fling, B.: Mobile Design and Development. O’Reilly Media Inc., Sebastopol (2009) 2. Stankovic, J.A.: Research directions for the internet of things. IEEE Internet Things J. 1, 3–9 (2014) 3. Jie, Y., Pei, J.Y., Jun, L., Yun, G., Wei, X.: Smart home system based on IOT technologies. In: International Conference on Computational and Information Sciences (ICCIS), pp. 1789– 1791 (2013) 4. Moser, K., Harder, J., Koo, S.G.M.: Internet of things in home automation and energy efficient smart home technologies. In: IEEE International Conference on Systems, Man, and Cybernetics (SMC), pp. 1260–1265 (2014) 5. Schilit, B., Adams, N., Want, R.: Context-Aware Computing Applications, pp. 85–90 (1994) 6. Yürür, O., Liu, C.H., Sheng, Z., Leung, V.C.M.: Context-awareness for mobile sensing: a survey and future directions, In: IEEE Communications Surveys & Tutorials, pp. 68–93 (2014) 7. da Costa, C.A., Yamin, A.C., Resin, C.F.R.: Toward a general software infrastructure for ubiquitous computing. IEEE Pervasive Comput. 7, 64–73 (2008) 8. Kindberg, T., Fox, A.: System software for ubiquitous computing. IEEE Pervasive Comput. 1, 70–81 (2002) 9. Salber, D., Dey, A.K., Abowd, G.D.: The context toolkit: aiding the development of context-enabled applications. In: SIGCHI Conference on Human Factors in Computing Systems, pp. 434–441 (1999) 10. Hristova, A., Bernardos, A.M., Casar, J.R.: Developing an ambient home care system: context toolkit-based design and implementation. In: Ninth International Conference on Parallel and Distributed Computing, Applications and Technologies, pp. 455–461 (2008) 11. Wang, W., De, S., Toenjes, R., Reetz, E., Moessner, K.: A comprehensive ontology for knowledge representation in the internet of things. In: IEEE International Conference on Trust, Security and Privacy in Computing and Communications, pp. 1793–1798 (2012) 12. Zafari, F., Papapanagiotou, I., Christidis, K.: Microlocation for internet-of-things-equipped smart buildings. IEEE Internet Things J. 3, 96–112 (2015) 13. Gerhard, F.: Context-aware systems: the ‘right’ information, at the ‘right’ time, in the ‘right’ place, in the ‘right’ way, to the ‘right’ person. In: International Working Conference on Advanced Visual Interfaces, 22–25 May, pp. 287–294 (2012)

Imperialist Competition Based Clustering Algorithm to Improve the Lifetime of Wireless Sensor Network Ali Shokouhi Rostami1, Marzieh Badkoobe1, Farahnaz Mohanna1, Ali Asghar Rahmani Hosseinabadi2(&), and Valentina Emilia Balas3

3

1 Department of Communication Engineering, University of Sistan and Baluchestan, Zahedan, Iran [email protected], [email protected], [email protected] 2 Young Researchers and Elite Club, Ayatollah Amoli Branch, Islamic Azad University, Amol, Iran [email protected] Aurel Vlaicu University, Bd. Revolutiei 77, 310130 Arad, Romania [email protected]

Abstract. In Wireless Sensor Network (WSN) nodes have limited energy and cannot be recharged. Clustering is one of the major approaches to optimize consumption of energy and data gathering. In these networks, clustering must be special to prolong network lifetime. In WSN, clustering has heuristic nature and belongs to NP-hard problems. In complex problems, search space is too big and grows exponentially. Because it takes too much time and cost, finding a deterministic optimized solution is difficult in such a short time. In this situation population-based algorithms are beneficial in finding optimum solutions. In this paper, a clustering algorithm is investigated and a novel idea, in line with the population-based algorithm, is presented. The proposed algorithm uses Imperialist Competition Algorithm (ICA) for the clustering of nodes. The results show that this algorithm postpones the dead time of nodes and prolongs network lifetime, compared to other discussed clustering algorithms. Keywords: Clustering
Imperialist Competition Algorithm
Lifetime
WSN

1 Introduction Wireless sensor network (WSN) consists of many tiny sensor nodes. These nodes are typically equipped with battery, processor, memory and radio to send and receive data, which will be deployed compactly in the area for a specific purpose. Having gathered the required information from the environment, sensor nodes that are connected with each other via wireless links process them, and then deliver them to Base Station (BS) that is responsible for gathering information [1, 2]. A wireless sensor node (node) is a small electronic device which has a limited energy resource, and generally the battery cannot be changed or recharged. Therefore lifetime of a sensor node is absolutely dependent on battery life [3, 4]. In WSN, most of © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_16

190

A.S. Rostami et al.

the energy is consumed in data transmission, therefore approaches that reduce transmission greatly save energy and increase network lifetime. Clustering is one of the most popular approaches used in reducing energy consumption. Clustering protocols, in comparison to direct protocols, not only decrease data transmission, but also, due to load balancing between nodes, reduce energy consumption [5, 6]. Selecting Cluster Heads (CH) and cluster formation procedures must make the clusters balanced, reduce the exchange of messages and keep the time complexity constant which is independent of the growth of the network. This selection process is very challenging [1, 7]. Clustering protocols proposed in the literature are, in general, classified into two groups: clustering protocols in homogeneous networks and clustering protocols in heterogeneous networks. The important clustering protocols in homogeneous networks are LEACH [8], extensions of LEACH [9, 10], PEGASIS [11], extensions of PEGASIS [12], TCCA [13], EEHC [14], HEED [15] and TEEN [16]. The important clustering protocols in heterogeneous networks are SEP [17], BSIDR [18], DEEC [19] and extensions of DEEC [20]. Clustering has a heuristic nature and belongs to NP-hard problems [9]. In heuristic problems, search space is too big and grows exponentially. In such cases, finding a deterministic answer needs too much time and cost, therefore it is difficult to find deterministic optimized answers in a short time. In this situation population-based algorithms are beneficial to determine optimum answers. Genetic Algorithm (GA) [21], Particle Swarm Optimization (PSO) [22], Gravitational Emulation Local Search Algorithm (GELS) [23–25], Gravitational Search Algorithm (GSA) [26–30] and Imperialism Competition Algorithm (ICA) are the popular population-based algorithms. These optimization protocols are applied to WSN to decrease the amount of data transmission and therefore prolong the network lifetime by finding the most suitable route in transmission and selecting optimal CHs [31]. These protocols are effective in many aspects like: optimal formation of clusters [32], selecting optimal nodes as cluster heads [10, 33], finding optimal number of clusters, and finding the optimal paths [34, 35]. In this paper, ICA is used for clustering. This protocol considers the conditions and limitations of WSN. In clustering topic, a protocol is needed that reduce data transmission and subsequently reduces energy consumption. This leads to an increase in network lifetime. In order to achieve these objectives, a new clustering protocol is proposed which first clusters all nodes using ICA and then in each cluster chooses the appropriate nodes as cluster heads. By using this protocol, more energy is remained in the nodes, and the lifetime of the network increases significantly.

2 Related Work Many algorithms in the field of energy efficiency have been proposed in the past and clustering is one of the most effective methods in this field. Following are a few of them.

Imperialist Competition Based Clustering Algorithm

2.1

191

Low-Energy Adaptive Clustering Hierarchy (LEACH)

LEACH [8] is one of the first and most popular energy efficient hierarchical clustering algorithms for WSN that is designed for reducing energy consumption. As the role of CH is rotated randomly and CHs directly forward their data to the BS with a one-hop route, this algorithm provides energy balancing. In LEACH formation of clusters and selection cluster heads, the CHs are completely distributed and there is no central coordination. CHs are selected randomly and all nodes have the same chance to be a cluster head. LEACH provides the following key areas of energy saving: • No overhead is wasted in deciding which node becomes cluster head as each node decides this independent of other nodes. • CDMA allows clusters to operate independently, as each cluster is assigned a different code. • Each node calculated the minimum transmission energy required to communicate with its cluster head and transmits with this optimized power level. 2.2

Stable Election Protocol (SEP)

SEP [17] is an extension of the LEACH algorithm. It is represented for heterogeneous networks. There are two types of node bases on initial energy and equipment: normal node and advance node. Advance nodes have more chances to be the cluster head. The other stages are similar to LEACH. This protocol improves lifetime and provides better energy balancing, compared to LEACH. 2.3

Distributed Energy Efficiency Clustering (DEEC)

DEEC [19] is used for both homogeneous and heterogeneous networks. Unlike LEACH and SEP, in this protocol cluster heads are not chosen randomly, but they are selected based on residual energy. According to this, in each round, nodes that have more energy, have a better chance of becoming the cluster head. All nodes must have global knowledge of the network to calculate the residual energy in each round. This adds a large overhead on the network.

3 Imperialist Competition Algorithm (ICA) To solve the optimization problems, different methods have been proposed such as Genetic Algorithm (GA) [36], Particle Swarm Optimization (PSO), and the Ant Colony Optimization (ACO). Imperialism Competition Algorithm (ICA) is a new evolutionary optimization method inspired by imperialism competition [37]. Like other evolutionary algorithms, ICA starts with an initial population called country [38]. There are two types of countries, colonies and imperialists. Some of the best countries in the population will be imperialists and the rest become colonies, which together forms the empires. Imperialistic competition takes place between these empires. During this competition, weak empires collapse and powerful ones take possession of their colonies. The power of an empire which is the counterpart of fitness function in GA, is inversely proportional to its cost.

192

A.S. Rostami et al.

After competition, the imperialists are selected and colonies are divided, then colonies move toward their imperialist. After dividing all colonies among imperialists and creating the initial empires, the competitions are started. Colonies start moving toward their relevant imperialist based on assimilation policy. Empires that cannot succeed in this competition and cannot increase their power (or at least prevent a power loss), are eliminated from the competition. This competition leads to a gradual increase in powerful empires and decrease in the weak ones. Eventually weak empires collapse and lose their power. Imperialistic competition converges to a state in which there exists only one empire and colonies have the same cost function value as the imperialist. In an N dimensional optimization problem, a country is a 1  N array. This array is defined as below: Country ¼ ½P1 ; P2 . . . PN  The cost of a country is found by evaluating the cost function f at the variables (P1, P2… PN), then CI ¼ f ðcountryi Þ ¼ f ðPi1 ; Pi2 . . . PiN Þ The algorithm starts with N initial countries called Npop, and Nimp best of them are chosen for imperialists. Ncol remaining countries are colonies that each belongs to an empire. Powerful empires have greater number of colonies and the weaker one have less number of colonies. To distribute the colonies among imperialists proportionally, the normalized cost of an imperialist is defined as follows: Cn ¼ maxi ci  cn

ð1Þ

where cn- is the cost of nth imperialist and Cn is its normalized cost. Imperialists began to improve their colonies and colonies move towards imperialists. This motion is shown in Fig. 1. In this movementɵ and x are random numbers with uniform distribution as illustrated in formula (2) and d is the distance between colony and the imperialist. x  Uð ; b  dÞ h  U ðc; cÞ

ð2Þ

where b is a number greater than 1. If b > 1, colonies and imperialist are closing together. c is an adjustment of deviation from the original direction. The values of b and c are arbitrary. The total power of each empire is determined by the sum of its power and average power of its colonies. TCn ¼ costðimperialistÞ þ n  meanfcostðcolonies of empiresÞg Where TCn is the total cost of n-th Empire and f is a positive number which is considered less than one.

Imperialist Competition Based Clustering Algorithm

193

The main steps of ICA are as follows: • Select some random point and initialize the empires. • Move colonies toward their relevant imperialist (Assimilation). • Randomly change the position of some colonies (revolution), If there is a colony with the imperialist. • Exchange the position of the colony and imperialist. • Until reaching similar empires. • Compute the total cost of all empires. • Pick the weakest colony from the weakest empire and hand it over to one of the empires (Imperialist competition). • Eliminate the powerless empires. • Exit if stop condition are satisfied, otherwise do further assimilation and continue.

Fig. 1. Colony moving toward imperialist [38]

4 The Proposed Algorithm In this work a WSN is considered as composed of n nodes which are the same and have two modes: Cluster Head Nodes (CH) and normal nodes. Since the energy consumption of CHs are greater than normal nodes; appropriate CHs are chosen in each cluster that needs less energy for transferring data to the base station and for receiving data from cluster members. This protocol is part of a fixed clustering algorithm and done with ICA. ICAAlgorithm consists of three phases: the set up phase, the cluster head selection phase, and the steady up phase. Set up phase is executed just once in the base station, whereas other phases are repeated in each round. In CH selection phase, appropriated CHs are chosen for each cluster. In the steady up phase, nodes gather data and then each node transfers its data to the corresponding CH’s. The CHs then transfer this received data to the base station.

194

4.1

A.S. Rostami et al.

Setup Phase

In this phase, nodes are clustered into kopt cluster using ICA. After nodes are located in their place, they send their located information to the base station. Then the BS, by using this information, calculates kopt. Nodes are clustered using ICA and the clusters are formed. BS broadcasts this information to all nodes. Location of nodes is given to the ICA as input parameters. In this case, finding the kopt point as the center of cluster is the objective. In so doing, sum of square distance of each cluster members to center of cluster, plus sum of square distance of center of clusters to BS, are minimized. 4.2

CH Selection Phase

Unlike the pervious phase which is performed only once in the BS, this phase is repeated in each round. After clusters are formed in the BS and cluster information is broadcasted to all the nodes, each node will realize who belongs to which cluster. Each node sends a message to its neighbors which are in the range of its transmission and thereby will inform its ID and cluster number. Each node receives its neighbor’s message, saves it in the table information of nodes which have similar cluster number to it. Then this table is sent to other cluster members. Using this table, every member node, can update its tables, thus all cluster members are identified. Most energy is consumed in transmission and transmission distance has direct and exponential direct in energy consumption. Center of cluster is center of gravity, too, for this reason if CH located around center of gravity, distance between all cluster members to CH will be approximately equal, hence energy consumption in transfer of data, will be almost same for all nodes. This causes better balance energy consumption of nodes and prevents preterm death. If CH is located around the center of gravity, it is optimized, however this condition is not sufficient for cluster head selection. As the numbers of nodes that are located around center, are limited, and if it was the only criterion, the energy of nodes adjacent to center is quickly finished and nodes will die soon. As a result, load balancing and energy balancing is destroyed. Hence, there are other criteria for selecting cluster heads. CHs consume more energy than member nodes, therefore the remaining energy of nodes must be noted in CH selection. Nodes which have more residual energy have higher priority than those which have less residual energy. Accordingly, in choosing CH, remained energy should be considered. Other criteria that are involved in selecting CH, is the distance between node and BS. The closer to the BS the CHs are placed, the lesser the energy used to transfer data. If this does not happen, the node that play role of CH, is remained as its role until its energy is finished end and node dies. This causes CHs to die sooner than members and the network balancing energy to be destroyed. The optimal status is when energy of all nodes is finished at the same time. Though, if the role of CH don’t replace, it will not happen.

Imperialist Competition Based Clustering Algorithm

4.3

195

Steady up Phase

This phase starts after CH selection phase and repeat in each round. To avoid collision and energy efficiency, TDMA scheduling is run and a time slot is allocated for each member node so that it can send its data to CH. To avoid wasting energy, the rest of time, they are inactive. CH receives data from all cluster members, then aggregate received data and send them to BS. When all CHs send their aggregated data to BS, the round is finished. Model of energy consumption is the same energy model proposed in [8] in order to achieve an acceptable. Signal-to-noise ratio (SNR) in transmitting an L bit message over a distance d, energy expanded by the radio is given by:
ETX ðL; d Þ ¼

L  Eelec þ L  efs  d 2 L  Eelec þ L  emp  d 4

if d  d0 if d d0

ð3Þ

Where Eelec is the energy dissipated per bit to run the transmitter or the receiver circuit, efs and emp depend on the transmitter amplifier model we use, and d the distance between the sender and the receiver. By equating the two expressions at d = d0, we have rffiffiffiffiffiffiffi efs d¼ emp

ð4Þ

To receive an L bit message the radio expends ERX ðLÞ ¼ L  Eelec

ð5Þ

5 Simulation Result For simulation of this algorithm, we have selected a rectangle area where 100 nodes are randomly distributed in the 100  100 area. The base station is located at the center (50, 50). The simulation is done with MATLAB. The parameters of simulation are listed in Table 1. Table 1. Simulation parameters Value 0.5 J 10pJ/bit/m2 0.0013pJ/bit/m4 50 nJ/bit 4000 bit 50 nJ/bit 87.7 m 3000

Parameters Initial energy Efs Emp EDA Packet length Eelec D0 Round number

196

A.S. Rostami et al.

In this simulation we suppose a network with assumptions as follows: • • • • •

All nodes are located randomly. All nodes are static and have limited energy. At the beginning energy of all nodes are the same. All nodes are aware of their position and their amount of remaining energy. All nodes can be both Cluster head and normal node.

For clustering of nodes, the first step is the calculating the optimal number of clusters. According to [8] and simulation parameters, it will be 75 m < dtoBS < 185 m. Therefore it is expected that the optimal number of clusters be in the range 1–6. Figures 2 and 3 show that by increasing the number of clusters, the average residual energy of the network is increased, death of nodes is delayed and network lifetime is

Fig. 2. Average residual energy in the various number of clusters

Fig. 3. Number of dead nodes in the various number of clusters

Imperialist Competition Based Clustering Algorithm

197

increased. In this experiment the optimal number of clusters is considered 6. ICA divides the network into 6 clusters so that the total distance of each cluster member to cluster center and the total distance from all cluster heads to the base station, is minimized. Figure 4 shows the formation of nodes in 6 clusters. Circles same in color are same cluster members and squares indicate center of each cluster (Fig. 5).

Fig. 4. Clustering nodes into 6 cluster using ICA

Fig. 5. ICA convergence rate

Simulation parameters for ICA are shown in Table 2. The result of this protocol is compared to discuss clustering algorithm. The results of these comparisons are shown in Figs. 6 and 7.

198

A.S. Rostami et al. Table 2. ICA simulation parameters Value 100 10 0.1 100 6

Parameter Number of Number of Revolution Iteration Number of

countries imperialist probability cluster

In this protocol, one of the cluster head selection parameter is proximity to the cluster center. The closer the nodes are to the center, the more their chance of being a cluster head is. Therefore the nearest nodes to center are the first priority to become cluster heads.

Fig. 6. Comparing numbers of alive nodes in each round

Figures 6 and 7 show in ICA, regardless of the percentage of nodes that are dead in early rounds, the remaining nodes almost die at the same time. As shown in Fig. 8, in early rounds ICA-Clustering have lesser average remaining energy than other algorithms. The reason is that some percentage of nodes die very quickly, but in other rounds ICA-Clustering have more remaining energy in each round and therefore the lifetime is improved than the others. ICA-Clustering is compared to

Imperialist Competition Based Clustering Algorithm

199

Fig. 7. Comparing number of dead nodes in each round

Fig. 8. Comparing number of average residual energy in each round

LEACH, SEP, DEEC and could obtain lifetime improvement about 40%, 36% and 26% respectively than them. The lifetime of this protocol is compared to some clustering algorithm shown in Fig. 9.

200

A.S. Rostami et al.

Fig. 9. Comparing lifetime of proposed algorithm and some clustering algorithm

6 Conclusion Wireless Sensor Network consists of a large number of tiny nodes that are distributed in the environment for specific purposes. Sensor nodes have limited energy and cannot be recharged. Algorithms used in WSN must both be energy efficient and improve network lifetime. Clustering is one of the approaches that is used for energy efficiency. In this paper we introduce a new fixed-clustering algorithm named ICA-Clustering. In this algorithm nodes are clustered with ICA in base station and then in those cluster nodes which are closer to the center of the cluster and have more residual energy. This improves its chance of being a cluster head. This algorithm, compared with other proposed clustering algorithms, leads to more energy efficiency and improves network lifetime.

References 1. Rabeay, J.M., Ammer, M.J., da Silva, J.L., Patel, D., Roundry, S.: PicoRadio supports ad hoc ultra-low power wireless networking. IEEE Comput. Mag. 33, 42–48 (2000) 2. Elahi, A., Hosseinabadi, A.R., Rostami, A.S.: Multi-hop fuzzy routing for wireless sensor network with mobile sink. Int. J. Sci. Eng. Res. 4(7), 2431–2439 (2013) 3. Kumarawadu, P., Dechene, D.J., Luccini, M., Sauer, A.: Algorithms for node clustering in wireless sensor networks: a survey, pp. 295–300, December 2008 4. Tavakkolai, H., Yadollahi, N., Yadollahi, M., Hosseinabadi, A.R., Rezaei, P., Kardgar, M.: Sensor selection wireless multimedia sensor network using gravitational search algorithm. Indian J. Sci. Technol. 8(14), 1–6 (2015) 5. Karenos, K., Kalogeraki, V., Krishnamurthy, S.: Cluster-based congestion control for sensor networks. ACM Trans. Sensor Netw. 4, 1–39 (2008) 6. Rostami, A.S., Bernety, H.M., Hosseinabadi, A.R.: A novel and optimized algorithm to select monitoring sensors by GSA. In: International Conference on Control, Instrumentation and Automation (ICCIA), pp. 829–834 (2011)

Imperialist Competition Based Clustering Algorithm

201

7. Elahi, A., Hosseinabadi, A.R., Rostami, A.S.: Improving news document clustering based on a hybrid similarity measurement. In: International Conference on Intelligent Computing and Intelligent Systems (ICIS), pp. 1–6 (2011) 8. Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H.: An application-specific protocol architecture for wireless microsensor networks. In: IEEE Tmns. Wireless Commun. pp. 660– 670, October 2002 9. Heinzelman, W.R., Chandrakasan, A., Balakrishnan, H.: An application-specific protocol architecture for wireless microsensor networks. In: IEEE Transactions on Wireless Communications, pp. 660–670, October 2002 10. Wang, A., Yang, D., Sun, D.: Clustering algorithm based on energy information and cluster heads expectation for wireless sensor networks. Comput. Electr. Eng. 38(3), 662–671 (2012) 11. Lindsey, S., Raghavendra, C.S.: PEGASIS: Powere-fficient Gathering in Sensor Information System. In: Proceedings IEEE Aerospace Conference, pp. 1125–1130, March 2002 12. Yueyang, L., Hong, J., Guangxin, Y.: An energy-efficient PEGASIS-based enhanced algorithm in wireless sensor networks, China Commun. Technol. Forum (2006) 13. Selvakennedy, S., Sinnappan, S.: An adaptive data dissemination strategy for wireless sensor networks, Int. J. Distrib. Sens. Netw., 3(1), 23–40 (2007) 14. Bandopadhya, S., Coyle, E.: An energy efficient hierarchical clustering algorithm for wireless sensor networks. In: Proceeding of IEEE INFOCOM, vol. 3, pp. 1713–1723, April 2003 15. Fahmy, S., Younis, O.: HEED: a hybrid energy-efficient distributed clustering approach for ad hoc sensor networks. IEEE Trans. Mobile Comput. 3(4), 366–379 (2004) 16. Manjeshwar, A., Agrawal, D.P.: TEEN: a protocol for enhanced efficiency in wireless sensor networks. In: The Proceedings of the 1st International Workshop on Parallel and Distributed Computing Issues in Wireless Networks and Mobile Computing, April 2001 17. Smaragdakis, G., Matta, I., Bestavros, A.: SEP: a Stable Election Protocol for clustered heterogeneous wireless sensor networks. In: Proceedings of the International Workshop on SANPA, pp. 251–261 (2004) 18. Varma, S., Nigam, N.: U.S. Tiwary, Base Station Heterogeneous Wireless Sensor Network using clustering, pp. 1–6 (2008) 19. Qing, L., Zhu, Q., Wang, M.: Design of a distributed energy-efficient clustering algorithm for heterogeneous wireless sensor networks. Comput. Commun. 29(12), 2230–2237 (2006) 20. Liu, Z., Zheng, Q., Xue, L., Guan, X.: A distributed energy-efficient clustering algorithm with improved coverage in wireless sensor networks. Future Gener. Comput. Syst. 28(05), 780–790 (2012) 21. Haupt, R.L., Haupt, S.E.: Practical Genetic Algorithms, 2nd edn. Wiley, Hoboken (2004) 22. Kennedy, J., Eberhart, R.: Particle swarm optimization. Proc. IEEE Int. 4, 1942–1948 (1995) 23. Hosseinabadi, A.R., Siar, H., Shamshirband, S., Shojafar, M., Nizam, M.H., Nasir, M.: Using the gravitational emulation local search algorithm to solve the multi-objective flexible dynamic job shop scheduling problem in Small and Medium Enterprises. Ann. Oper. Res. 229(1), 451–474 (2015). Springer 24. Rostami, A.S., Mohanna, F., Keshavarz, H., Hosseinabadi, A.R.: Solving multiple traveling salesman problem using the gravitational emulation local search algorithm. Appl. Math. Inf. Sci. 9(2), 699–709 (2015) 25. Hosseinabadi, A.R., Kardgar, M., Shojafar, M., Shamshirband, S., Abraham, A.: GELS-GA: hybrid metaheuristic algorithm for solving multiple travelling salesman problem. In: International Conference on Intelligent Systems Design and Applications (ISDA), pp. 76–81 (2014)

202

A.S. Rostami et al.

26. Hosseinabadi, A.R., Yazdanpanah, M., Rostami, A.S.: A new search algorithm for solving symmetric traveling salesman problem based on gravity. World Appl. Sci. J. 16(10), 1387– 1392 (2012) 27. Hosseinabadi, A.R., Farahabadi, A.B., Rostami, M.S., Lateran, A.F.: Presentation of a new and beneficial method through problem solving timing of open shop by random algorithm gravitational emulation local search. Int. J. Comput. Sci. 10(1), 745–752 (2013) 28. Hosseinabadi, A.R., Ghaleh, M.R., Hashemi, S.E.: Application of modified gravitational search algorithm to solve the problem of teaching hidden Markov model. Int. J. Comput. Sci. 10(3), 1–8 (2013) 29. H. Tavakkolai, A. R. Hosseinabadi, M. Yadollahi, T. Mohammadpour, “Using Gravitational Search Algorithm for in Advance Reservation of Resources in Solving the Scheduling Problem of Works in Workflow Workshop Environment”, Indian Journal of Science and Technology, Vol. 8(11), 1–16, June 2015 30. Hosseinabadi, A.R., Kardgar, M., Shojafar, M., Shamshirband, S., Abraham, A.: Gravitational search algorithm to solve open vehicle routing problem. In: 6th International Conference on Innovations in Bio-Inspired Computing and Applications (IBICA 2015), Chapter Advances in Intelligent Systems and Computing, Kochi, India, pp. 93–103. Springer (2016) 31. Shijun, H., Yanyan, D., Zhou, R., Zhao, S.: A clustering routing for energy balance of WSN based on genetic algorithm. In: International Conference on Future Computer Support Education, IERI Procedia, vol. 2, pp. 788–793 (2012) 32. Shahvandi, L.K., Teshnehlab, M., Haroonabadi, A.: A novel clustering in wireless sensor networks used by imperialist competitive algorithm. Int. J. Adv. Eng. Sci. Technol. 8(2), 276–280 (2011) 33. Bayraklı, S., Zafer Erdogan, S.: Genetic algorithm based energy efficient clusters (GABEEC) in wireless sensor networks. In: The 3rd International Conference on Ambient Systems, Networks and Technologies, vol. 10, pp. 247–254 (2012) 34. MurtalaZungeru, A., MinnAng, L., PhooiSeng, K.: Classical and swarm intelligence based routing protocols for wireless sensor networks: a survey and comparison. J. Netw. Comput. Appl. 35, 1508–1536 (2012) 35. Bore Gowda, S.B., Puttamadappa, C., Mruthyunjaya, H.S., Babu, N.V.: Sector based multi-hop clustering protocol for wireless sensor networks. Int. J. Comput. Appl. 43(13), 33– 38 (2012) 36. Shojafar, M., Kardgar, M., Hosseinabadi, A.R., Shamshirband, S., Abraham, A.: TETS: a genetic-based scheduler in cloud computing to decrease energy and makespan. In: 15th International Conference on Hybrid Intelligent Systems (HIS 2015), Chapter Advances in Intelligent Systems and Computing 420, Seoul, South Korea, vol. 420, pp. 103–115. Springer (2016) 37. Shamshirband, S., Shojafar, M., Hosseinabadi, A.R., Abraham, A.: OVRP_ICA: an imperialist-based optimization algorithm for the open vehicle routing problem. In: International Conference on Hybrid Artificial Intelligence Systems (HAIS), vol. 9121, pp. 221–233. Springer, LNCS (2015) 38. Atashpaz-Gargari, E., Lucas, C.: Imperialist competitive algorithm: an algorithm for optimization inspired by imperialistic competition. IEEE Congr. Evol. Comput. 7, 4661– 4666 (2007)

Wireless Sensor Networks Relay Node Deployment for Oil Tanks Monitoring Ola E. Elnaggar, Rabie A. Ramadan(B) , and Magda B. Fayek Cairo University, Giza, Egypt ola [email protected], [email protected], [email protected]

Abstract. Oil tanks monitoring is an important problem in the field of oil production. Tanks might be distributed in large areas and dedicated people might be assigned for each tank to monitor it in terms of its temperature, the surrounding area, and even the level of oil inside. Wireless Sensor Networks (WSNs) could be one of the best solutions to this problem. The problem of these tanks is that they are far from each other and they might have obstacles among them that block the sensors’ signals. At the same time, WSN consists of many low-cost nodes with limited power and communication. In addition, these sensors are either randomly deployed in the monitored field or deterministically installed in specific places. In both cases, gaps may appear in the network leaving uncovered areas/tanks. In this paper it is improved the connectivity of the network used for oil tanks monitoring by adding minimum number of relay nodes to it. To solve the problem of oil tanks monitoring by using WSN are used the following algorithms: the Divided Network Area Algorithm (DNAA) by dividing the network area to small squares and connecting theses small areas; the Best Path Algorithm (BPA) by making the network connected when selecting the best path for all sensors and guarantee that all nodes are connected; and the Adjustable Communication Range with Best Path Algorithm (ACR-BPA) working with both the communication ranges and the network paths. With different problem settings, the proposed algorithms are examined and compared to a state-of-art greedy algorithm.

1

Introduction

Wireless Sensor Networks (WSNs) are taking huge attention due to its wide range of application in the last few years. Some of these applications are air pollution monitoring, health care monitoring, forest fire detection, Oil & Gas Remote Monitoring applications for example pipeline monitoring [9] and tank level monitoring [10], etc. WSN consists of large number of wireless sensor nodes, which are deployed in particular area to measure certain phenomenon such as temperature, sound, pressure, etc. These sensors send their measured data to a central processing unit, which collects all the data and develops the decision. Each sensor has its sensing rang, communication rang and limited power. Sensing rank is the monitored area around the sensor. Communication rank is the minimum distance between two sensors to send the data between them. c Springer International Publishing AG 2018  V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8 17

204

O.E. Elnaggar et al.

Due to the limitations of the wireless sensors prolonging their lifetime becomes the important attribute in WSN. Another problem in WSN is the gaps between sensors due to the random deployment of the sensors. Random deployment of sensor nodes most of the time makes WSN disconnected. To solve this problem we need to add some relay nodes to the network to make it connected. In order to achieve better communication and maximize the network lifetime, relay nodes are deployed in the network using an efficient deployment algorithm. One of the most important applications of WSNs is monitoring oil tanks. Oil tanks are containers available in many shapes: vertical and horizontal cylinder; open top, close top and flat bottom. Sensors could be deployed in/on the tanks to measure certain features such as tank’s temperature, oil level, and surrounding environment. However, with large number of tanks, the connectivity of these sensors might be a problem. To solve this problem we need to deploy a minimum number of relay nodes on the disconnected areas in the network to make the network connected. At the same time, WSN lifetime is another issue that has to be considered during the relay deployment which makes the problem more complex. The paper presents three algorithms to solve the monitoring oil tanks problem. The first algorithm, entitled Divided Network Area Algorithm (DNAA), in which it divides the network area into grid of cells, could be in a form of squares. Relay nodes will be added to the center of these cells for their purpose is to enhance sensors connectivity. The second algorithm, entitled Best Path Algorithm (BPA), exploits the concept of best path to the centralized node (sink node/base station). It computes the best path for each sensor node and tries to add the relay nodes on these paths. The last algorithm is named Adjustable Communication range with Best Path Algorithm (ACR-BPA). This algorithm is a modification to BPA with the use of one of the sensors features which is the adjustable communications ranges. The remainder of the paper is organized as follows. Section 2 states some of the most related work to the work done in this paper. Section 3 presents the problem definition. The algorithms involved to solve the oil monitoring problem are shown in Sect. 4. Section 5 presents the simulation and the comparison of the algorithms. Conclusion of the paper and the future work is presented in the final section.

2

Related Work

The scientific literature includes papers that consider the connectivity and the energy consumption in WSNs [1–6]. In [1], the authors considered deploying of few additional nodes possible to reconnect a disconnected network. In [2,7], the main goal is to fully enhance the connectivity of the wireless sensor networks by adding an available set of relay nodes to the network after going through some stages called levels. In the first level, the authors divided the network area into certain numbers of equal regions and represented each region by a relay in its center. In the

Wireless Sensor Networks Relay Node Deployment for Oil Tanks Monitoring

205

second level, the best K relays’ locations are chosen by solving a semi-definite programming (SDP) optimization. Through the third level, it iteratively refines the solution by dividing each obtained relay’s region into a number of smaller regions and repeating the same procedure. In [3], a relay sensor placement algorithm to maintain the connectivity is proposed. They formulated this problem into a network optimization problem, named Steiner Minimum Tree with Minimum Number of Steiner Point (SMT-MSP). This study restricts the transmission power of each sensor to smallest value, then add relay nodes to guarantee the connectivity. In [4], three heuristic algorithms are proposed for achieving connectivity of a randomly deployment ad hoc wireless networks. This work connects the network with minimum number of relay nodes with maximum utility from a given number of the additional nodes for the disconnected network. Another work proposed in [6] where the authors studied single-tiered constrained relay node placement problems, under both the connectivity requirement and the survivability requirement. In this paper, we study relay node deployment problem in WSN. Oil tanks monitoring is considered as a realistic case study application. In oil tank problem we need to make all sensors on each tank connected to the base station by adding minimum number of relay nodes in the optimal places to prolong the network lifetime.

3

Oil Tanks Monitoring Problem

The Oil & Gas industry is one of the most prevalent industries for the application of Wireless Sensor Technology. WSN solutions for the oil, gas and power generation reducing maintenance costs, lowering the energy consumption, minimizing downtime, improving equipment performance, enhancing safety and centralize controls. Wireless technology has the potential to be beneficial in many regards. Eliminating the need for cables can contribute to reduced installation and operating costs; it enables installations in remote areas and it allows for cost efficient. Examples on Oil & Gas Remote Monitoring applications are Pipeline Integrity Monitoring and Tank Level Monitoring. In this paper, we focus on Tank Level Monitoring problem using WSN. Oil tanks are containers holding oil with open top, close top and Oil Pumping Units (OPU). Most of the Oil Pumping Units (OPU) are manually monitored (Fig. 1). This oil-pumping system use a high power-consuming process and is incapable of OPU’s structural health monitoring. In Tank Level Monitoring with WSN, the condition of the oil storage tanks can be monitored using sensors: level sensor, temperature sensor and gas sensor. These sensors are fixed inside the oil storage tanks. The sensor output is given to the sink node. Based on the condition of the oil storage tanks and the oil pumping motor is controlled. The motivations of using WSN in Tank Level Monitoring are: the special nature of oil exploration (the majority of oil pumping units are spread over barren hills, mountains and deserts) and the existing oil-pumping systems with manual control.

206

O.E. Elnaggar et al.

Fig. 1. Wireless sensor network on area L × L.

Manual control systems have three disadvantages: (1) The OPU administrator must go to the oil field frequently to see the OPU status regardless the weather and the place. (2) Power consumption for OPU is huge during the oilpumping process. (3) It is hard to fix the monitoring problem because it depends on people. For these reasons, WSN is used to monitor oil tank levels. Let consider a wireless sensor network in a two dimension space. All sensor nodes (terminal nodes) need to send all information about oil level on a certain tank to a sink node. The connectivity of these sensors depends on the position of the sensors and their communication ranges. As can be seen in Fig. 2, as an example, all sensors need to send all data to the Sink Node (SN). Sensor S2 can send to SN without a problem because S2 is directly connected to SN but for sensor S1, some relay nodes need to be added for the purpose of connectivity. Therefore, the work in this paper is step towards solving this problem using minimum number of relay nodes and saving the overall network lifetime.

4

Algorithms for Solving the Oil Tanks Monitoring Problem

In this section, three algorithms are proposed to solve the oil tanks monitoring problem. The algorithms try to add some relay nodes to the WSN to make it connected with minimum number of relay nodes and try to prolong the network lifetime as well. The algorithms are: Divided Network Area Algorithm (DNAA), Best Path Algorithm (BPA) and Best Path Algorithm with Ant Colony Optimization (BPA-ACO). The following subsection introduces the Ant Colony Optimization algorithm that will be further used in the BPA-ACO algorithm.

Wireless Sensor Networks Relay Node Deployment for Oil Tanks Monitoring

4.1

207

Ant Colony Optimization

Ant Colony Optimization (ACO) is a class of algorithms. The first algorithm, Ant System, was initially proposed by Colorni, Dorigo and Maniezzo [11–13]. The base of ACO is to simulate the real behavior of ants in nature. The ant colony provides indirect communication with the help of ant’s pheromone. Pheromones are chemical substances which attract other ants searching for food. The attractiveness of a given path depends on the quantity of pheromones detected by an ant. The quantity of pheromone is governed by some rules and depends on the attractiveness of the route. The use of attractive route ensures that the ant exudes more pheromones on its way back and so that the path is much attractive for other ants. The evaporation of pheromones is time dependent. When the way is no longer used, pheromones are evaporated and the ants begin to use other paths. ACO steps are as follows. First Step: For first sensor, martificial ant start with random number from 1 to maximum number of communication levels (in this paper there are 6 levels). Second Step: Each ant builds a solution by adding one communication level after the other until it reaches the last sensor. The selection of the next communication range depends on certain probability. The probability pik of transition of a virtual ant from the node i to the node k is given by formula 1; τi - indicates the attractiveness of transition in the past, ηi - adds to transition attractiveness for ants, ni- set of nodes connected to point i, without the last visited point before i, β, α- system dependent parameters. (η α + τiβ ) pik =  i β (ηni α + τni )

(1)

Third Step: Pheromone update, virtual ant is using the same reverse path as the path to the food source based on its internal memory, but in opposite order and without cycles. After elimination of the cycles, the ant puts the pheromone on the edges of reverse path according to formula (2); τij (t) is the value of pheromone in step t, Δτ is the value by ants saved pheromones in step t. Values Δτ can be constant or they can be changed depends on solution quality. τij (t + 1) = ρτij (t) + Δτ (t)

(2)

Fourth Step: At last, the pheromones on the edges are evaporated. The evaporation helps to find the shortest path and provides that no other path will be assessed as the shortest as given in Eq. (3); ρ is a user-defined parameter called evaporation coefficient. (3) τij (t + 1) = (1 − ρ)τij (t) In our problem, a group of artificial ants searches solutions by starting each ant with random communication level for first sensor and completing the path until last one finishes depending on certain probability. This operation is inspired from the Theorem 1 given in [8]:

208

O.E. Elnaggar et al.

Theorem 1. To maximize the lifetime of the WSN, for the sensor nodes Ni and Nj, the power level assignment should satisfy xi ≥ xj for i < j. The high communication range is given high probability when the sensor is far away from the sink node and low probability is given to the nearest sensor to the sink node. When all ants reach to the end of the path, the pheromone gets updated depending on the path length and the consumed power. Therefore, the path covering the pipe with law power consumption has high pheromone value. ACO will stop after reaches the maximum number of iterations and returns the best path. 4.2

Greedy Algorithm

In [8] is solved the oil tanks monitoring problem using the Greedy algorithm. The greedy algorithm starts with certain number of sensor nodes are deployed with certain position (the position of sensor nodes are on the top of oil tanks). The greedy algorithm includes the following three phases: Initial Graph Phase: Find the initial graph G(Nt , E(Lt , Rt )) where Nt is the terminal nodes with location Lt and communication range Rt . When the distance between two sensors is less than the communication range, the two sensors are connecting. In this phase, no relay nodes are added. Constructing Delaunay Phase: Construct Delaunay by using the terminal nodes. The construction of the Delaunay is illustrated as follows. Let S be the set of the points in the two dimension space. The Voronoi diagram of S, denoted as Vol(S) which is decomposed into Voronoi cells {Va: a ∈ S} defined as: V a = {x ∈ R2 : |x − a| ≤ |x − b|, ∀b ∈ S}. The dual of the Voronoi diagram is the Delaunay triangulation Del (S ). Del (S ) is geometrically realized as a triangulation of the convex hull of S. After constructing Delaunay, the algorithm calculates the length of the three edges of each triangle. If the length of the edge is not larger than the transmission range Rt, then connect it. Triangle Type Phase: After Phase 1, the algorithm divides the Delaunay triangles into three types. In Type 1, the length of all edges of the triangle is larger than Rt and smaller than 2Rt. In Type 2, the longest edge of the triangle is at most 4Rt, while the shortest edge is larger than Rt and at most 2Rt. The properties of triangles different from Types 1 and 2 are defined as Type 3. For the triangles of Type 1. The algorithm places one relay node to connect five nodes that are formed by three adjacent triangles. Second, it places one relay node to connect four nodes that are formed by two adjacent triangles. Third, it adds one relay node to connect three nodes of one triangle. For the triangles of Type 2, it tries to place two relay nodes to connect three nodes of one triangle. For the triangles of Type 3, it adds relay nodes to connect the nearest disconnected nodes pair along the edge of the triangle. See [8] for other details on the greedy algorithm.

Wireless Sensor Networks Relay Node Deployment for Oil Tanks Monitoring

4.3

209

Divided Network Area Algorithm

The Divided Network Area Algorithm (DNAA) starts with a certain number of sensor nodes (N ) deployed on the top of the oil tanks. All sensor nodes start with the same communication level Rt and the same power level. Sensor nodes assume to send its own data about the oil level, temperature, pressure, and any other features from the oil tank to the ink node (SN). DNAA includes the following phases: Dividing Area Phase: Divide the network area to squares with edges equal to the communication rang Rt . On each square which include at least one sensor, add one relay node at the center of it. DNAA didn’t add any relay nodes in squares without sensor nodes. Each two sensors with distance less than Rt are connected. In Fig. 2 are ten sensors. Assume sensor number 1, s1, is the sink node. All sensors need to send the data to s1. After DNAA had divided the area to squares, it adds relay nodes at the center of each square that has at least one sensor in it, see s12, s13, s15, etc. then, it draws a line (line with dark blue) between the two sensors that have distance d between them where d ≤ Rt (s12 and s3). Connecting Sensor Nodes Phase: Make the disconnected node connected with the nearest node with adding relay nodes like sensor number 6 in Fig. 3 or without relay node if the distance between the two sensors is less than Rt . For instance, from Fig. 2, to connect s1 with s2 DNAA adds a relay node s18 and the same for s6 and s15, it adds s19. Relay Nodes Connected Phase: Make the main relay nodes (relay nodes added in the center of the square) connected by adding other relay node(s) on the shortest path between the two of them as shown in Fig. 3. All of the sub-graphs need to be connected (Fig. 3). DNAA connects subgraphs: s1, s2, s11, s18 with sub-graphs: s13, s14, s4, s5 and some relay nodes s20:s23. Repeat this phase until all of the sub-graphs are connected.

Fig. 2. (a) Before deploying the relay nodes, (b) after adding relay nodes at the center of the square.

210

O.E. Elnaggar et al.

Fig. 3. Making the disconnected nodes connected (left) and the disconnected relay nodes connected (right).

4.4

Best Path Algorithm

BPA starts with the same conditions for greedy algorithm and DNAA. It starts with a certain number of sensor nodes (N) are deployed on the top of the oil tanks. All sensor nodes are assumed to have the same communication rang Rt and the same power level. Sink node collect all data from all sensor nodes in the WSN. The BPA consist of the following phases. Connection Without Relay Node Phase: In this phase, each two sensors with distance less than Rt are assumed connected. Connection With Relay Node Phase: In this phase BPA adds relay nodes to make the network connected. BPA first computes all paths from each sensor to the sink node. BPA calls the path as the best path if the number of relay nodes added to make the path connected was minimum number than others. To do this, BPA gives each link in the path a weight. The weight of a disconnected link is the needed number of relay nodes to make it connected. Every connected link has a weight equal to zero. The path weight is equal the sum of all links’ weight, see Fig. 4. As in Fig. 4, all sensors need to send its data the sink node (s1 is the sink node in this example). The links with green color mean that the two sensor are connected (the distance between them is less than Rt ) with weight w = 0. Links with red color means that the distance between the two sensors is greater than Rt ; in this case, the algorithm must add some relay nodes to make the link connected. Links with w = 3 means that they need three relay nodes to the link connected. For instance, sensor 4 can send its data to the sink node through many paths. For example, Path 1 goes through nodes 4, 9, 2, and 1 with weight = weightlink (4:9) + weightlink (9:2) + weightlink (2:1) which is equal to 1 + 0 + 0 = 1. Path 2 goes through nodes 4, 3, and 1 with weight equals to 0+3 =3. Sensor 4 will send its data through path 1 because it has the minimum weight. In the next step link (4:9) will be with w = 0 because it will be connected. The algorithm will repeat step two until all sensors will be connected.

Wireless Sensor Networks Relay Node Deployment for Oil Tanks Monitoring

211

Fig. 4. Path weight in BPA. (Color figure online)

4.5

Adjustable Communication Rang with Best Path Algorithm

ACR-BPA assumes sensors with adjustable communication range R1, R2 . . . R6 as in [8] instead of using one communication range. The algorithm starts with a certain number of sensor nodes deployed on the top of oil tanks with certain positions. All sensor nodes are assumed to have the same communication power and adjustable communication range. In the first step, the algorithm selects the best path for all sensors while the second step consists of two parts: Part A: Choose the communication range (RC) of the sensors depending on the distance between them. For example, in Fig. 4, if the distance d between two sensors is less than or equal to R6 (the maximum communication range), ACRBP doesn’t add any relay node. When: R5 < d ≤ R6, RC for the two sensors is R6; R4 < d ≤ R5, RC for the two sensors is R5 . . . d ≤ R1, RC for the two sensors is R1. Part B: Add the relay nodes to the link when d > Rmax. To deploy relay nodes in this step ACR-BP uses Ant Colony Optimization (ACO) deployment algorithm.

5

Numerical Experiments and Discussions

In this section, the results for using WSN in Level Tank Monitoring problem will be illustrated. The next subsections show the effect of using the proposed algorithms on the network lifetime in case of different network settings. Experiment 1: Comparison among the four algorithms in terms the required numbers of relay nodes. In this experiment, there are certain number of oil tanks are established randomly on square area with length L, L = 100 m. One sensor node is deployed on the top of each tank. The experiment studies the effect of using the four algorithms to make the wireless sensor nodes connected.

212

O.E. Elnaggar et al.

Fig. 5. The number of relay nodes added in the four algorithms in square area with L = 100 m (left) and L = 500 m (right).

The wireless sensor nodes, n, (where n is the number of oil tanks) changes from n = 10 to n = 100 sensors. All sensors as assumed to have the same power level. Sensors communication ranges as assumed the same for all algorithms except for ACR-BPA in which it utilizes different communication ranges in its operation. From the results in Fig. 5, DNAA used the maximum number of relay nodes even more than greedy algorithm. ACR-BPA seems to be the best in terms of the required number of the relay nodes added for the network to be connected. ACR-BPA uses the minimum number of relay nodes compared with the rest of the algorithms. These results apply to all different networks with different number of nodes. Therefore, the algorithms could be ranked from the best to the worst in terms of the deployed number of relays as ACR-BPA, BPA, Greedy, and then DNAA. In fact, the implementation of DNAA on a real environment could be somehow difficult due to the special nature of oil exploration and oil drilling, the majority of oil pumping units (OPU) are spread over barren hills, mountains and deserts. Experiment 2: Comparison among the proposed algorithms in terms of the required numbers with large monitored area. In this experiment, again the number of oil tanks are changed from 10 to 100; however, the monitored area in this set of experiments are considered as a square of side length L = 500 m. This large area certainly affects the required number of relay nodes. Therefore, the purpose of this set of experiments is to examine the behavior of the proposed algorithms as well as the greedy algorithm. As can be seen in Fig. 5, DNAA still requires the maximum number of relay nodes; on the other hand, ACR-BPA used the minimum number of relay nodes. At the same time the behavior of the greedy algorithm and BPA is almost similar in terms of the number of the relay nodes they use. ACR-BPA enhances the results from Greedy algorithm with almost 23.2%, and BPA acts much better than the greedy algorithm with almost 0.009%.

Wireless Sensor Networks Relay Node Deployment for Oil Tanks Monitoring

213

Fig. 6. The effect of adding relay nodes on the network lifetime in the four algorithms on square area with L = 100 m (left) and L = 500 m (right).

Experiment 3: Network lifetime under the operation of the proposed algorithms. In this set of experiments, we examine the performance of the proposed algorithms compared to the greedy algorithm in terms of network lifetime. The area of oil tanks is assumed a square shape L×L where L = 100 m. A sensor is deployed on the top of each oil tank. The term network lifetime mean, in this context, either the first node dies or a relay nodes dies. The network lifetime is computed as the number of the round before any or both of these conditions occurs. As shown in Fig. 6, the number of oil tanks as well as the number of sensors change from 10 to a 100 on a square area of 100 × 100 m. Oil tanks are deployed randomly in all cases. The Fig. 6 shows that ACR-BPA gives the best lifetime over all of the other algorithms. For instance, at 60 oil tanks, the network still operated for 26 rounds. In addition, DNAA algorithm comes in the second rank after the ACR-BPA in which at 60 oil tanks, it keeps the network operate for 24 rounds. At the same time, BPA and the greedy algorithms have the same performance for almost 12 rounds with 50 oil tanks are deployed. The results is confirmed in Fig. 6 even when the monitored area is increased to a square with side length L = 500 m. Experiment 4: Comparison among the proposed algorithms in terms of the required number of relay nodes and network lifetime with monitored area of equilateral triangle shape. Here the performance of the algorithms are examined when equilateral triangle area are used. The network topology in this case is assumed to differ from the square monitored area. There is a certain number of oil tanks are established randomly on equilateral triangle area with edge L = 200 m. One sensor node is deployed on the top of each tank. The set of experiments illustrate the effect of the network shape on the network lifetime and the number of relay nodes. The number of oil tanks increases from n = 10 tanks to n = 100. DNAA can’t implement in this experiment because it requires the monitored areas to be divided area into square shapes which is not valid in our case. From the results shown in Fig. 7, the change in the shape of the network area (oil tank area) causes a change in the required number of relay nodes and

214

O.E. Elnaggar et al.

Fig. 7. The number of relay nodes added (left) and effect of adding relay nodes on the network lifetime (right) into the four algorithms in equilateral triangle area with L = 200 m

the network lifetime. However, ACR-BPA still over-performing other algorithms in terms of required number of relay nodes and network lifetime. At the same time, BPA and the greedy algorithm performing almost similar to each other especially with the large number of oil tanks. For instance after 60 oil tanks are deployed in the monitored field, their curves are overlapping as can be seen in Fig. 7.

6

Conclusion and Future Work

In this paper, the problem of oil tanks monitoring is solved using three proposed algorithms. The algorithms are compared to each other and with an existing greedy algorithm. ACR-BPA algorithm based on ACO seems to have the best performance over the compared ones. However, ACR-BPA is suitable only for nodes with multilevel communication ranges. Therefore, BPA could be suitable for nodes with fixed communication range and ACR-BPA will be the best for network with nodes having variable communication range. The conducted extensive set of experiments show that the ACR-BPA and BPA are much better than the greedy algorithm in terms of the number of deployed relays and network lifetime. For our future work, we would like to examine the performance of the proposed algorithm on network with different topologies and monitored areas with different shapes. Acknowledgements. The study was conducted under the auspices of the IEEE-CIS Interdisciplinary Emergent Technologies task force.

References 1. Li, N., Hou, J.C.: Improving connectivity of wireless ad hoc networks. In: Proceedings of the Second Annual International Conference on Mobile and Ubiquitous Systems: Networking and Services (MobiQuitous 2005), pp. 314–324 (2005)

Wireless Sensor Networks Relay Node Deployment for Oil Tanks Monitoring

215

2. Ibrahim, A.S., Seddik, K.G., Ray, K.J.: Improving connectivity via relays deployment in wireless sensor networks. IEEE GLOBECOM, pp. 1159–1163 (2007) 3. Cheng, X., Du, D.Z., Wang, L., Xu, B.: Relay sensor placement in wireless sensor networks. Wirel. Netw. 14(3), 347–355 (2008) 4. Koskinen, H., Karvo, J., Apilo, O.: On Improving Connectivity in Static Ad Hoc Networks by Adding Nodes. Med-Hoc-Net (2005) 5. Hou, Y.T., Shi, V.Y., Sherali, H.D., Midkiff, S.F.: On energy provisioning and relay node for wireless sensor network. IEEE Trans. Wirel. Commun. 4, 2579–2590 (2005) 6. Misra, S., Hong, S.D., Xue, G., Tang, J.: Constrained Relay Node Placement in Wireless Sensor Networks to Meet Connectivity and Survivability Requirements (2008) 7. Chang, J.-H., Jan, R.-H.: An efficient relay sensor placing algorithm for connectivity in wireless sensor network. In: Embedded and Ubiquitous Computing. Lecture Notes in Computer Science, Vol. 4096, pp. 874–883 (2011) 8. Guo, Y., et al.: Sensor placement for lifetime maximization in monitoring oil pipelines, In: Proceedings of the 1st ACM/IEEE International Conference on Cyber-Physical Systems, pp. 61–68 (2010) 9. Elnaggar, O.E., Ramadan, R.A., Fakry, M.B.: WSN in monitoring oil pipelines using ACO and GA. In: The 5th International Conference on Sustainable Energy Information Technology (SEIT-2015), vol. 52, pp. 1198–1205 (2015) 10. Barani, R., Jeyashmi, V.: Oil well monitoring and control based on wireless sensor network using atmega 2560 controller. Int. J. Comput. Sci. Commun. Netw. 3, 341–346 (2013) 11. Colorni, A., Dorigo, M., Maniezzo, V.: Distributed optimization by ant colonies. In: Proceedings of ECAL 1991, European Conference on Artificial Life. Elsevier Publishing, Amsterdam (1991) 12. Dorigo, M., Maniezzo, V., Colorni, A.: The ant system: an autocatalytic optimizing process, Technical report TR91-016. Politecnico di Milano (1991) 13. Dorigo, M.: Optimization, learning and natural algorithms. Ph.D. thesis. Politecnico di Milano, Milano (1992)

A Smart Way to Improve the Printing Capability of Operating System Adeel Ahmed1, Muhammad Arif1(&), Abdul Rasheed Rizwan2, Muhammad Jabbar1, and Zaheer Ahmed1 1

Department of Computer Science, University of Gujrat, Gujrat, Pakistan [email protected], {m.arif,m.jabbar,zaheer}@uog.edu.pk 2 Department of Computer Science, Virtual University of Pakistan, Lahore, Pakistan [email protected]

Abstract. Operating system is the core of computer science and task scheduling is the major topic in this domain. Priority base scheduling is always remains hot topic in the domain of operating systems. In Priority base printing always higher priority given to those printings jobs which are processed more quickly rather than the lower priority base printings jobs. In this way low priority jobs are delayed again and again. In this research detail study is conducted regarding the scheduling algorithms in operating system. Priority base printings become bottlenecks who are assigned lowest priority. They may have to wait till their turn may be up to mid night in large departments. Now we are going to discuss a specific algorithm which can solve above issues in a smart way. A new method is proposed to solve the priority scheduling problems. This method is very handy for those users who are assigned a higher priority authority in printing mechanism. It also allows the lowest priority base printing to print data. In this solution we are not going to ignore the highest printing jobs. The highest priority base printing jobs will be continued in the same way. Our main objective is to maintain a balance between the high priority and low priority printing jobs without suffering from continues delay. Keywords: Priority printing
Printer scheduling
Printing capability




Spooler




Algorithm




Smart

1 Introduction Priority base scheduling is always remains hot topic in the domain of operating systems. Printing Architecture (PA) is a major component of any Operating System. PA is consists on printer driver and the application which sends the prints to printer [1]. Operating system is the core of computer science and task scheduling is the major topic in this domain. Printer drivers are attached with spool simultaneous peripheral operations on-line used to add printing job to buffer. Spooling provides best behavior to such I/O which has different data rates in receiving and sending mechanisms. Spooler is a kind of queue which takes input from rear and out at front part of the queue. The spooler sends data to target printer’s I/O port which can be in serial, network, USB, wireless © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_18

A Smart Way to Improve the Printing Capability of Operating System

217

communication channel and parallel [2]. In Priority base printing always higher priority given to those printings jobs which are processed more quickly rather than the lower priority base printings jobs. In this way low priority jobs are delayed again and again. In this research detail study is conducted regarding the scheduling algorithms in operating system. Priority base printings become bottlenecks who are assigned lowest priority. They may have to wait till their turn may be up to mid night in large departments. Every printing job has some operating to complete its printing processes. For e.g., user sends printing command from an application which include direct printing, networks printing and internet base printing, after sending a command from the application it enters to the spooler (Simulates Peripheral Operation online) [3]. It also allows the lowest priority base printing to print data. In this solution we are not going to ignore the highest printing jobs. The highest priority base printing jobs will be continued in the same way. Our main objective is to maintain a balance between the high priority and low priority printing jobs without suffering from continues delay. Now we are going to discuss a specific algorithm which can solve above issues in a smart way. A new method is proposed to solve the priority scheduling problems. This method is very handy for those users who are assigned a higher priority authority in printing mechanism [4]. It also allows the lowest priority base printing to print data. In this solution we are not going to ignore the highest printing jobs. The highest priority base printing jobs will be continued in the same way. Our main objective is to maintain a balance between the high priority and low priority printing jobs without suffering from continues delay. The spooler performs specific processing on the sent job and passes to the printer driver. The printer driver translates the job according to the bitmap of the printer. Printing job to local or remote computer is performed by printer provider. Printer provider is responsible for print queue, functions and Printing Priority Queue (PPQ) [5]. PPQ performs high priority job before the low priority for printing. We can set the priority of individual documents in the general tab, use the priority slider (Windows OS) low priority is 1 and highest priority is 99.

2 Related Works Priority queue is different from normal queue. The normal queue data structure is base on FIFO. But priority queue is a kind of box which receives different priority jobs and sent the highest priority out to the calling function. Here is a conceptual picture of a priority queue is shown in the Fig. 1. Figure 1 shows how that printing jobs are inserted in to the queue and the highest priority job is taken out at first. It also shown that the lowest priority jobs are remained in the queue waiting for their turn. Printing jobs are sent to priority queue to sort their positions; different kinds of method are used for sorting their priorities just as array representation, link list representation and heap sort etc. [4]. All these data structure have computation time to sort the requested prints. If a printer is already working on hundreds request and so much busy and in that position while a printer is going to print documents at this position it may receive another request and it have to compute sort on all prints, and this situation may exists for long time. This can stop the working of all prints [6]. We can see before sorting priority queue jobs, this job can enter the priority

218

A. Ahmed et al.

Fig. 1. Priority queue [2]

queue in different order. After entering the jobs into priority queue an algorithm is applied to sort them. Absolutely this algorithm takes time to sort them. Figure 2 illustrates the before sorting view.

Fig. 2. Before sorting overview

Starvation can be caused in PPQ. Just e.g., there 1 to 99 priorities for users according to their work in a origination. 99 have the highest priority and 1 has the lowest priority in case of Microsoft OS. From 99 to let’s say 20 users are already in printing but remaining 19 to 1 may b in starvation. It can be from 1 to 19 has urgent information to give his boss but he is the victim of starvation. So just PPQ is not so better in such kinds of case for printing jobs. Figures 3 and 4 illustrates the priority Queue.

Fig. 3. Priority queue [7, 8]

A Smart Way to Improve the Printing Capability of Operating System

219

Fig. 4. Priority queue enqueue and dequeue [13]

In Fig. 4 it is shown that enque and dequqe functions are takinf the highest priority job first and then move to the lowest priority jobs. Assigning different priority level for different kinds of groups with high and low priority level of queue is applied to Microsoft Windows server edition [5]. According to Microsoft to use the best way of priority level printing, create multiple logical printer of the target printer. Each priority level should be different to other group of priority level. For example users in Group A should be able to connect with printer with highest priority just as priority number one. And Group no. 2 should access the printer with the priority level two and further so on [9, 10, 16]. The method printing not only support LAN but also the WAN in client systems [11]. This will become a huge network so we have to manage it in a smart way. The method of printing first include to identify the server method to receive and print the jobs [16]. We can customize printing priority level for others groups. Click on start button and click on Open Printers and Faxes. Right click on target printer, click on selected printer Properties, and then choose Advanced tab. Now click on Priority tab, click the up or down arrows, and then click OK. Here you can set priority level from 1 to 99. “99” is the highest priority level and “1” is lowest priority level in printing [17]. Select the appropriate priority level and click OK. Now to add new logical printer so to add a new logical printer click on Add Printer to add another logical printer for the same existing physical printer. Now again we have to set its priority level according to our feet needs. Allow the low level users to use the first logical printer and medium level users allow them to use second logical

220

A. Ahmed et al.

printer, following the way we set high level user for the highest logical printer. When we need to open Printer and Faxes dialog box, we have to click on start button and next click on Printer and Faxes [12].

3 Problem Statement Microsoft does not allow it to manage lowest priority jobs automatically. Microsoft Server provides manually setting to overcome this issue. In Microsoft environment we have to change the priority of documents which are waiting to get printer. Even the selected job’s priority changing does not affect other priority jobs, yet it is not a good solution for large business environments. It is also happen in Microsoft Server edition the lowest priority base printing is fixed up to a fix time of night. The lowest printing priority is selected at midnight for printings. The time setting is hard set so in time of free printer, the time set priority documents are not allowed to printer to use the printer resource. So this solution to the problem is not so much efficient. Higher priority should in top ascending order numeric numbers and lower priority job should rear part of ascending order in numeric numbers.

4 Results and Discussions Queue is a linear data structure where an item is removed from top end (Front) and new item is inserted at back end (Rear). This is shown in the Fig. 5.

Fig. 5. Queue insertion and deletion

Figure 5 shows the insertion and deletion of the jobs in the queue box. This is a common fashion which is used in normal kind of queue and data structure. But we are going to use this queue data structure in a different way with the help of linked list. Linked list allow to add and remove a node in anywhere of the list. We don’t have to shift up and down of other part in linked list to remove or to add a node as in array data structure. Let’s suppose the following is a linked list queue. In above figure different priority numbers are shown in Fig. 6, let’s suppose these are the priority numbers of printing commands. And we need to send command of

A Smart Way to Improve the Printing Capability of Operating System

221

Fig. 6. Priority numbers

priority number two, which is recently entered into queue. But due to its priority we have to send it first to printer. Here we need to remove this node from the linked list. Now we can see according to its data structure nature we should send command of top node to printer. We also not are going to sort it according to its priority, because we don’t want to use CPU cycle here due to efficiency issue. After removing the target node, now see in Fig. 7 [13].

Fig. 7. Target node removal

After knowing all kinds of issues in PPQ lets solve these issues in a smart way [14]. Before going in detail of our algorithm first take overview the flow chart of our algorithm. This describes the working of algorithm in better understanding [15, 16]. 4.1

Proposed Methodology (Boss Algorithm)

Boss Algorithm is operating section of our algorithm. It should receive the print command before going to spooler. In standard scenario printing command works as [Application > Spooler > Printer]. But in BA this order should [Application > BA > Spooler > Printer]. When BA will receive a printing command from an application it should check the priority of the printing job. If the priority is one, it should directly send to Ready Priority Queue (RPQ). In other case if the sent priority print is not one it should send to Printing Priority Queue (PPQ) [17]. The Boss Algorithm is depicted in the Fig. 8. The pseudo code of the algorithm is given below. If (priority == 1) { Move to the RPQ Spooler Printing } Else { Move to PPQ RPQ Spooler Printing }

222

A. Ahmed et al.

Fig. 8. Boss algorithm

4.2

Printing Priority Queue (PPQ)

PPQ should receive all kinds of prints which have not priority of one. Suppose we have following prints which are recently sent to PPQ. Before going in detail of PPQ, first let discuss its structure. In Fig. 9 we can see the different prints have arrived in PPQ, different color boxes telling us the different prints in PPQ. Above white color numeric numbers are telling us the priority numbers which are 3, 5, 13, 8, 7 up to 9. Below numbers of the figure such as 0  001, 0  002 so on up to 0  00n. These numbers are addresses of the prints which are in PPQ in a link list point of view. PPQ data structure should program with link list. It is because we need add and delete operation, we know add and delete operations are much fast in link list against any other data structure. When a print enters into PPQ, its priority should be written and it address should also append with the print. Other words we can say that two kinds of information should have with every print. The Fig. 6 has different priorities without having the highest number priority [18]. When PPQ will receive a print having priority number 2, it should send before other prints to RPQ with the same priority number. We can see in Fig. 6 there is one print which has the priority number two. It should directly send to RPQ without further processing. It is because we also need a difference in RPQ which is receiving only priority number one. If we change the priority up to one in PPQ then it will be difficult for RPQ which job should send to printer first. So we should send print from PPR to RPQ in priority number two. When RPQ send command to printer which having the priority number two, RPQ should signal to PPQ I am serving your request and now you should rescan your printing jobs again. This process should do on each process which having the priorities number two in RPQ. PPQ is responsible for following processes.

A Smart Way to Improve the Printing Capability of Operating System

223

I. Send the print to RPQ which has priority number 2 without further processing. II. Increase the priority up to two, which has low priority than 2. III. After sending print to RPQ, PPQ is responsible to delete the node and join nodes. Now we take an example to understand PPQ. Now again see the following picture of Fig. 9.

Fig. 9. Prints arrival in the queue

After finishing all jobs which have the priority number 1, RPQ should signal to PPQ to send it printing jobs. So PPQ should send the priority number two printing job to RPQ, after it PPQ should delete the node after sending RPQ.

Fig. 10. Print job is sending to Ready Priority Queue

Figure 10 is showing that priority number 2 print job is sending to Ready Priority Queue (RPQ) without scanning and increasing priority number. After sending the print job, now PPQ should delete and join the sent printing job.

224

A. Ahmed et al.

Fig. 11. Delete and join node

Figure 11 is showing after sending, deleting and joining the node process. Next PPQ should wait and receive all prints commands while PPQ does not receive the signal of scanning. After scanning the signal of scanning PPQ should scan again all prints and increase the priority with number one, if there is again a printing command which the priority number two, it should directly send to RPQ without further processing. In Fig. 11 we can see there are two print commands which have priority number 3, mean on one scanning its priority should 2.

Fig. 12. Increasing the priority number

Now we can see the difference in Figs. 11 and 12 after scanning and increasing the priority number. Now we can see there are two prints commands which have priority level two. Now PPQ should send these two prints commands to RPQ, and after it PPQ should delete these nodes.

A Smart Way to Improve the Printing Capability of Operating System

225

Fig. 13. PPQ after deleting and joining two nodes

Figure 13 is showing PPQ after deleting and joining two nodes. Now PPQ should again wait signal from RPQ for further scanning, increasing priority, sending print jobs, deleting and joining nodes. 4.3

Ready Priority Queue (RPQ)

RPQ should attach to spooler for sending prints commands to printer. RPQ should have two kinds of priority type prints number one, and number two. Priority prints which have priority one, should send to printer before the printing having the priority number two. RPQ can differentiate easily that which print should send to printer first. When priority having number one should not in RPQ then priority two should send to printer for printing in batch sequence [19]. If printer is busy in priority two batch sequence and a new printing command inter in RPQ, after completing the prints of the under process printing then next turn should priority number one. When RPQ is working on priority number two, it should signal to PPQ for further scanning and increasing the priority which it has in its queue. This situation may happen there is one print RPQ and PPQ have many other prints jobs. There is no such priority exists which level increase to two for RPQ in just one scan. On empty queue of RPQ is should signal again and again to send print command to it. Now let’s take an example to understand RPQ.

Fig. 14. Two types of priorities are in RPQ

In Fig. 14 we can see two kinds of priorities are in RPQ, the priority number one came directly from BA after checking its priority. The priority number two is come from PPQ. This number of priority might be up to level two. In PPQ if the priority is two it should directly send to RPQ without increasing and extra processing over it.

226

A. Ahmed et al.

In Fig. 14 the right side command first came into RPQ so it should send to spooler first, second from the right side second last print command came into RPQ so it should send to spooler at position number two. Now next is priority number two which is in third number from right side. Because we have more prints which have priority number one, so the number two priority should not send to printer. The next print is also the priority number one which is at fifth number from left to right in Fig. 14.

Fig. 15. Way of numbering

Figure 15 is showing us the way of numbering in which prints should send to spooler for printing. The new coming prints should treat in FCFS fashion if they have the same priority, the only difference is that priority number one and priority number two. On serving priority number two RPQ should signal to PPQ to rescan the priority numbers and increase the numbers according to its appropriate sequence. 4.4

Join Node (JN)

Join Node is sub program part of PPQ, JN should perform two kinds of work. It should first remove the current sent print job node, after it, it should join the deleted node left and right nodes.

Fig. 16. Complete picture of join node

A Smart Way to Improve the Printing Capability of Operating System

227

Figure 16 is showing a complete picture of JN, JN should work Delete and Join Node. When a print should send to RPQ, PPQ should call next Join Node function. The first part of the Join Node should delete the sent print command node, after deleting the node next JN should call its next function to join node [20]. While working in Deleting and joining the nodes PPQ should wait and not to scan the nodes. When JN return its control to PPQ, then PPQ should precede its next work.

5 Conclusion and Future Work BOSS algorithm is best way for printing all kinds of priority level. It provides turn to every level of priority and it also gives first priority to the highest level priority users. It never let down or star-vat the low priority users. Low priority level users not have to wait their print in mid night. There is no need to add virtual printer and extra overhead for server and low priority level users. BA is complex in sense handling all kinds of priorities in same time. Change the priority of the queued jobs in every search setup the over head of CPU. In this way this proposed solution can be implemented in huighly demanded environment where printing capabilities are lacked. In future we are planning to implement this scenario in well developed organization and gain the benefits of newly designed technique. This will be more beneficial for the print and press industries to groom their business. This idea can be utilized for priority based process scheduling mechanism.

References 1. Macauley, R.M., Martin, T.S., Vachon, G.C., Cosgrove, P.A., Sasson, S.J., Wilson, E.D.: Distributed Printing Social Network. Google Patents (2015) 2. Shaw, L.F., Teng, C.-C., Sykes, K.W., Endres, R.E.: Device Independent Spooling in a Print Architecture. Google Patents (1998) 3. Laurent, M.S., Onischke, M., Kuindersma, M., Krishnammagaru, D., Stairs, J., Noreikis, K.: System and Method for Releasing Print Jobs Based on Location Information. Google Patents (2015) 4. Sanders, R.E., Frazier, D.P.: System and Method for Batch Printing High-Volume Electronic Documents from a Network. Google Patents (2010) 5. Gould, K.V.W., Small, G.J., Galipeau, S.R., Kirkland, C.J.: Methods and Systems for Data Prioritization. Google Patents (2014) 6. Zhang, D., Dechev, D.: A lock-free priority queue design based on multi-dimensional linked lists. IEEE Trans. Parallel Distrib. Syst. 27, 613–626 (2016) 7. Gruber, J., Träff, J.L., Wimmer, M.: Benchmarking Concurrent Priority Queues: Performance of k-LSM and Related Data Structures. arXiv preprint arXiv:1603.05047 (2016) 8. Lenharth, A., Nguyen, D., Pingali, K.: Priority queues are not good concurrent priority schedulers. In: European Conference on Parallel Processing, pp. 209–221 (2015) 9. Kinoshita, K., Higashizaki, Y.: Remote e-mail printing. Google Patents (2011) 10. Nan, X., He, Y., Guan, L.: Optimal resource allocation for multimedia cloud based on queuing model. In: 2011 IEEE 13th international workshop on Multimedia signal processing (MMSP), pp. 1–6 (2011)

228

A. Ahmed et al.

11. Abaev, P., Gaidamaka, Y., Samouylov, K.E.: Queuing model for loss-based overload control in a SIP server using a hysteretic technique. In: Internet of Things, Smart Spaces, and Next Generation Networking. Springer, pp. 371–378 (2012) 12. Nakayama, Y., Takewa, Y., Sumikura, H., Yamanami, M., Matsui, Y., Oie, T., et al.: In-body tissue-engineered aortic valve (Biovalve type VII) architecture based on 3D printer molding. J. Biomed. Mater. Res. Part B: Appl. Biomater. 103, 1–11 (2015) 13. Lansing, S., Pantelias, N., Vu, Y., Gomez, F.J.: System and Method for Dropping Lower Priority Packets that are Slated for Wireless Transmission. Google Patents (2011) 14. Al Hanbali, A., Alvarez, E., Heijden, M.: Approximations for the waiting time distribution in an M/G/c priority queue (2013) 15. Stanford, D.A., Taylor, P., Ziedins, I.: Waiting time distributions in the accumulating priority queue. Queueing Syst. 77, 297–330 (2014) 16. Singh, C.J., Jain, M., Kumar, B.: MX/G/1 queuing model with state dependent arrival and second optional vacation. Int. J. Math. in Oper. Res. 4, 78–96 (2012) 17. de Souza, R.M., Morabito, R., Chiyoshi, F.Y., Iannoni, A.P.: Incorporating priorities for waiting customers in the hypercube queuing model with application to an emergency medical service system in Brazil. Eur. J. Oper. Res. 242, 274–285 (2015) 18. Takagi, Y., Nakagawa, M.: Printing System, Printing Method, Print Server, Control Method, and Computer-Readable Medium. Google Patents (2012) 19. Alistarh, D., Kopinsky, J., Li, J., Shavit, N.: The SprayList: a scalable relaxed priority queue. ACM SIGPLAN Not. 50, 11–20 (2015) 20. Kim, C., Choi, S.: A Study on Efficient Join Mechanism Using Streaming-Service-Time in Mobile P2P Environment (2014)

A Comparative Study for Ontology and Software Design Patterns Zaheer Ahmed, Muhammad Arif(&), Muhammad Sami Ullah, Adeel Ahmed, and Muhammad Jabbar Department of Computer Science, University of Gujrat, Gujrat, Pakistan {zaheer,m.arif,msamiullah,adeel.ahmed, jabbar.ahmed}@uog.edu.pk

Abstract. Ontology design patterns have been extended from software design patterns for knowledge acquirement in semantic web. The main concern of ontology design patterns is to how concepts, relations and axioms are established in a ontology using ontological elements. Ontology design pattern provide the solution of quality modeling for ontologies. In user prospective the presentation of ODP play a central role for the reusability of ontologies. This research determine the improvement areas in the presentation of content ODPs. Improvement in presentation can ultimately improve the understandability of a pattern from user perspective. Our objective is to analyze the template of different software engineering patterns (SEP) and ODP. On the basis of this analysis we suggest possible changes in current template and pattern presentation. It also includes determining the most important information about patterns which can help an ontology engineer in selecting an appropriate pattern. Presentation of design patterns is related to issues such as reuse, guidance and communication. Our main goal is to evaluate the current patterns presentation. The evaluation is focused on the analysis of current patterns. The ontology design pattern templates were compared with existing templates of other patterns for determine the improvement areas. The template of an ODP consists of many parts, the first question is to identify the most important and vital information concerning the design patterns. This information would help an ontology engineer to select an appropriate design pattern for the required ontology. The second question is about the users who work with ontology design patterns. Generally, users are divided into two categories; novice and expert ontology engineers. Novice users are the end-users who use design patterns to implement in the ontologies. Expert ontology engineers are those who actually develop ontology design patterns. Each category of user has its own information requirement regarding design patterns. Keywords: Ontology design pattern
Templates pattern
Software engineering pattern




Content ontology design

1 Introduction ODPs are semantic patterns that provide quality modeling solutions for ontologies. ODPs play an important role in learning and teaching ontology engineering [1]. They facilitate the automatic and semi-automatic ontologies construction and provide a base © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_19

230

Z. Ahmed et al.

for creating ontologies in different domains [2]. In user prospective the reusability of ontology depends on presentation of ODP. So far only a small catalogue of patterns exist which is available online at the ontology design pattern portal. In this portal, ODPs are described using a template with a set of headings that should be filled out when entering a new pattern. The template defines a standard way for constructing new patterns. There are possibilities to discuss modeling issues, review and suggest changes in patterns [4]. In computer and information science ontology is defined as a “formal, explicit specification of a shared conceptualization” [3]. One of the main problem areas is reusability of ontologies. The existing ontologies are available at online ontology repositories which provide guidelines to ontology users. Due to unfamiliar logical structure the existing ontologies provide limited support. Must be learned the good practices form literature. This problem is solved by implementing common solution as we learn in software engineering [4]. The patterns facilitate and to some extent automate the construction of ontologies. The development of patterns in the ontology field is very popular as that in software engineering. The patterns are defined for reuse and aim at facilitating the construction process very much like the way it is done in software engineering or architectural planning of buildings [5]. The purpose of design patterns is to solve the design problems. The patterns provide a useful way for handling the problems of reusability in a development process. In SE the common practices to build software through design and architecture patterns. This practice also follow in ontology engineering [2]. Ontology design patterns provide modeling solutions of ontologies design problems. They provide a base for creating ontologies in different domains. Patterns are also used for evaluation of ontologies [4]. Ontology design patterns (ODPs) are of several types. They are divided into 6 families; Content patterns, Structural Patterns, Presentation patterns, Correspondence Patterns, Lexico-Syntactic Patterns and Reasoning Patterns [6]. This thesis deals with the presentation of content ontology design pattern. It describes the design issues of the presentation of ontology design patterns. There may be certain information that an ontology user need to understand a pattern but it is not available in the description, our next task will be to examine the missing information in the current ontology design pattern templates. Finally, based on the results of above questions, we will made suggestions for the potential improvement in the current templates of ODP presentation.

2 Related Work Knowledge reuse is common way to improve the quality of work artifacts. Pattern is a common way to improve reusability. There are other ways to support reusability, i.e. in object oriented programming the concept of program components related to the reusability of design. To achieve reusability, number of design technique available in object oriented software development model for more reusable building block. An obstacle for reuse methodologies is the lack of motivation among developers. Before

A Comparative Study for Ontology and Software Design Patterns

231

starting the process, the developer needs to establish a reuse library which requires extra efforts. Reuse process is divided in to two steps Design for reuse and Design by reuse. To expedite design by reuse, first the design for reuse process must be established [5, 7]. Reusability is applied at different levels. In software engineering, reusability can be applied at following three levels [5]: Requirements Reuse: It deals with the models of the domain or the generic model of the requirement domain. Design Reuse: It deals with the models, data structures and algorithms. Software Component Reuse: It deals with reuse of software classes and editable source code. The development process of ontology is time consuming and more effort required form developer. Reuse of an ontology from an ontological engineering perspective can be hard. This is even more when there are large ontologies to be reused [8]. Ontology Library: For ontologies to be grouped and organized so that they may be reused further and for ontology integration, maintenance, mapping and versioning, an important tool known as ontology library systems was developed. These systems must fulfill all ontological reuse needs and must be easily accessible [9]. Ontology Matching: Ontology matching is a process of judging correspondence between semantically related ontologies, solving the problem of semantic heterogeneity and can be used in ontology merging, query answering, data translation etc. So ontology matching facilitates interoperability between matched ontologies [10]. A pattern is something re-occurring that can be applied from one time to another and also from one application to another. These concepts are in use in our daily life and also in our professional life. We use old solution as patterns. We search patterns in our surroundings that can be useful. In reuse the pattern are best practices that provide the good design [5, 11]. Currently in the computer science field the most common and popular patterns are software patterns. Most of the software development projects, where applying functional or object oriented design are conducted using patterns. The patterns are used for increasing reusability, product quality and for managing complexity of the system development process. According to the phase of the development process where they are used these patterns are divided into different kinds [5]. The most common categories are the following: • • • •

Analysis Patterns [5, 12]. Architecture Patterns [5, 13]. Design Patterns (see [11, 14–17]). Programming Language Idioms (see [3, 13]).

232

Z. Ahmed et al.

Analysis patterns are used to describe the conceptual structures of business processes for the different types of business domains like how to transform these processes into software [3, 12, 18]. An overall structuring principle is used while constructing a viable software architecture. Architectural patterns are considered as templates for solid software architectures [3]. Design pattern describe the common solution of overall design problem in certain context through the depiction object and class communication. Design pattern provide the description of participation classes, their instance, their roles and interaction, distribution of responsibility. One object oriented design problem address by the certain design pattern that also provided sample code to describe the partial solution of the problem [14]. The lowest level of patterns is represented by idioms. The implementation of particular design issue is dealt with by the idioms. The aspects of both, the design and implementation, are treated by them [13]. 2.1

Problem Statement

The ontology design patterns (ODPs) have divided in grouped of six different families: Structural ODPs, Content ODPs, Reasoning ODPs, Presentation ODPs, Correspondence ODPs and Lexico Syntactic ODPs [6]. Graphically the types of patterns are represented as (Fig. 1):

Codesolution: Ontology Design patterns

Reasoning ODP

Structural ODP

Architecture ODP

Content ODP

Logical ODP

Lexio-Syntatic ODP

Name ODP

Presentation ODP

Annotation ODP

Reengineering ODP

SchemaReengineering ODP

LogicalMacro ODP

Correspondence ODP

Aligment ODP

Refactoring ODP

Transformation ODP

Fig. 1. Group of ontology design pattern [19]

In this paper, our focus of research will be the Content ODPs. The comparison of ODPs is limited to software patterns and data model patterns because these are the most used and well known patterns. There are many ontology languages available for development but we will only focus on OWL ontologies [19].

A Comparative Study for Ontology and Software Design Patterns

233

This research guide the improvement areas in the presentation of content ODPs. Improvement in presentation can ultimately improve the understandability of a pattern from user perspective. The focus of research is to analyze different content ODPs and provide some possible recommendation in current templates of the ODP presentation. It also includes determining the most important information about patterns which can help an ontology engineer in selecting an appropriate pattern. Presentation of design patterns is related to issues such as reuse, guidance and communication. Our main goal is to evaluate the current patterns presentation. The evaluation is focused on the analysis of current patterns. The template of an ontology design pattern consists of many parts, the first question is to identify the most important and vital information concerning the design patterns. This information would help an ontology engineer to select an appropriate design pattern for the required ontology. The second question is about the users who work with ontology design patterns. There may be certain information that an ontology user need to understand a pattern but it is not available in the description, our next task will be to examine the missing information in the current ontology design pattern templates.

3 Methodologies There are many types of patterns but we have limited our comparison to those of software engineering and data models patterns, since these are the well know and common use patterns. Software patterns have been compared to ontologies by Devedzic in one of his article [16] where the author argued that there is a significant overlap between the two concepts and that it is the aim, while generality and practical usage of these concepts differ. The concepts of patterns and ontologies have some common goals, e.g. sharing and reusing of knowledge. Both these concepts necessitate hierarchies of concepts, relationships, vocabularies and constraints. Moreover, both of them can be seen as using an object-oriented paradigm [3]. First, we analyze what current templates exist in other fields and then compare them to ontology design pattern templates to analyze the difference. In our study, evaluation was done on the results of the evaluation of different patterns. Templates of the patterns were compared to identify the difference and similarities in their presentation. Each part of the templates was studied with respect to its objective and the content provided in that part [19]. A template of a pattern is a standard way of representing a pattern. In a broad sense, a pattern template has four important elements. These elements are: Name, Problem, Solution and Consequences [3, 11, 19]. The different kinds of pattern templates are given below with their description.

234

3.1

Z. Ahmed et al.

Design Pattern Template

This template proposed by Guarino [11] in their book “Design pattern Element of Reusable Object-Oriented Software” [11]. Table 1 shows the different parts of the template and their description (Table 2).

Table 1. Design pattern template [11, 19] Design pattern template Elements Description Pattern name and Name is a short summary of the pattern. There are several DP, we classification need a way to classified them in a family. The section classification refers these families of design pattern Intent It is a brief description that explain the following. How the design pattern work? What is the main goal of the pattern and what are the particular design issues or problem solve by the pattern Also known as Another name of the pattern, If the pattern has other name Motivation It has a scenario that describes a design problem and explain the class and object structures in the pattern describe the problem solution. The problem solution will facilitate you to understand the more abstract description of the pattern Applicability This section discus the situations in which the design pattern applicable Structure It illustrates a detailed specification of the structural aspects of the pattern. It includes a graphical representation of the classes in the pattern using the notation of OMT (Object Modeling Technique). This section also has interaction diagrams to illustrate sequences of requests and collaboration diagram for description of collaboration between objects Participants This section describes the different parts of the pattern and their relation. In design pattern the participants are classes and/or objects Collaborations This section describes how the participants collaborate to carry out their responsibilities Implementation Implementation gives guidelines for implementing the pattern. It gives hints and techniques which one should be aware before implementing the pattern. For example if there are language specific issues Sample code Sample code is a code fragment that illustrates how you might implement the pattern in a programming language Known uses Known Uses is the examples of the use of the pattern in real systems. It includes a minimum of two examples from different domains Related patterns Related patterns are described, i.e. what are the closely related patterns to this given pattern? What are important differences? With which other patterns should this one be used?

A Comparative Study for Ontology and Software Design Patterns

235

Table 2. Builder design pattern [19] Builder design pattern Elements Pattern Name and Classification Intent

Also Known As Motivation Applicability

Structure Participants Collaborations

Implementation Sample code

Known uses

Related patterns

Description Builder, creational patterns To split the construction of the complex object from its representation so that[same construction process can create different representations Fig. 2 When the builder pattern are used • The algorithm for creating a complex object should be independent of the parts that make up the object and how they’re assembled • The construction process must allow different representations for the object that’s constructed Fig. 3 Builder, ConcreteBuilder, Director, Product The given interaction diagram describe how Builder and Director cooperate with a client Fig. 4 Fig. 5 /Abstract Builder class abstract class TextConverter{ abstract void convertCharacter(char c); abstract void convertParagraph(); } . . . . public static void main(String args[]){ Client client = new Client(); Document doc = new Document(); client.createASCIIText(doc); system.out.println(“This is an example of Builder Pattern”); } } The RTF converter application is from ET++. Its text building block uses a builder to process text stored in the RTF format Abstract factory

236

Z. Ahmed et al.

Fig. 2. Builder design pattern

Fig. 3. Builder design pattern

Fig. 4. Sequence diagram

A Comparative Study for Ontology and Software Design Patterns

237

Fig. 5. Class diagram

3.2

Analysis Pattern Templates

Given below is the Analysis Pattern template described by Fernandez and Liu in their article “The Account Analysis Pattern” [20]. This template is also described in the book “Pattern-Oriented Software Architecture” [21] (Tables 3 and 4). Table 3. Analysis pattern template [19, 21] Analysis pattern template Elements Description Pattern name It describes the name for referring to the pattern and also other names if the pattern has another name Intent It is a short statement that answers many questions like what does that design pattern do? What is the main goal of the pattern and what are the particular design issues or problems solved by the pattern Example Provides a real world example, which shows an existing problem and exemplifies the need of the pattern Context The section context is a fundamental component of a pattern. It provides an indication of the applicability of a pattern Problem It defines the recurring problem that is solved by the general solution. Problem is a fundamental component of a pattern because it is the reason for the pattern. The problem which is addressed by the pattern is described in this section Solution The solution details the participating entities in the solution, the collaborations between them and their behavior Example resolved Example Resolved gives the solution of the given example Know uses Know Uses is the example of the use of the pattern in a real system. It includes a minimum of two examples from different domains Consequeses It details the benefits that a pattern can offer and any possible restrictions Related patterns Related patterns are described what are the closely related patterns to this given pattern? What are important differences? With which other patterns should this one be used?

238

Z. Ahmed et al. Table 4. Account analysis pattern [19, 22]

Account analysis Elements Pattern name Intent

Example

Content

Problem

Solution

Example resolved

Know uses

Consequneses

Related patterns

pattern template Description Account Analysis Pattern The Account pattern keeps track of accounts of customers in institutions. These customers can perform transactions of different types against the accounts Consider a banking institution, where customers have accounts of different types, e.g., checking, savings, loan, mortgage, etc. For the convenience of their customers, the bank may have several branches or offices located in different places There are many institutions, e.g., banks, libraries, clubs, and others, that need to provide their customers or members with convenient ways to handle financial obligations, charge meals, buy articles, reserve and use materials, etc. Without the concept of account users need to carry large amounts of cash, may have trouble reserving items to buy or borrow, and would have serious problems sending funds to remote places Start from class Account and add relevant entities; in this case customers, cards, and transactions. Build an institution hierarchy describing the branches of the institution and relate accounts to the branches Fig. 6 An example for Bank accounts is shown in Figure. The classes contained in the model include Bank, BranchOffice, Account, CheckingAccount, Customer, BankCard, TransactionSet (TXSet), and Transaction, with their obvious meanings. Class TXSet collects all the transactions for a user on his account for a given period of time. There are, of course, other types of accounts Fig. 7 The following are examples of uses of this pattern: • Banks, where customers have financial accounts of different types • Libraries, where patrons can borrow books and tapes • Manufacturing accounts, where materials are charged The pattern has the following advantages: • It is clear that this model provides an effective description of the needs and can be used to drive the design and implementation of the software system. Not using a similar model would result in code that is hard to extend and probably incorrect • One can easily add other use cases: freeze account and activate/deactivate account The liabilities of this pattern come from the fact that to limit the size of the pattern and to make it more generic we have left out: Different types of customers. Each variety of customers could be handled in a special way. Accountability pattern

A Comparative Study for Ontology and Software Design Patterns

239

Fig. 6. Class diagram

Fig. 7. Class diagram

3.3

Architecture Pattern Templates

This template was described by Ayodele Oluyomi in his article “Patterns and Protocols for Agent-Oriented Software Development” for the Agent internal ArchitectureStructure Patterns [22] (Table 5). Example The pattern description has been slightly abbreviated for readability issues (Table 6).

240

Z. Ahmed et al. Table 5. Architecture pattern template [19, 21]

Architecture Pattern Template Elements Description Name A brief summary of the pattern Problem It defines the problem which a pattern can solve Context Different kinds the circumstances in which a pattern can be applied Forces It contains description of various forces and constraints that can affect the desired objectives Solution This section describes the different part of the pattern and their relation Known uses Know Uses are examples of the use of the pattern in real system. We include minimum two examples from different domains Result This section description of possible effects on the initial context when the context solution is applied and also the resulting advantages and disadvantages Related patterns are described; What are the closely related patterns to this Related pattern given pattern? What are important differences? With which other patterns should this one be used? Table 6. Agent as delegate pattern [19, 22] Elements Name Classification Problem Context Forces

Solution

Known uses Result context

Related pattern

Description Agent as delegate Multiagent system architecture-definitional How should the role of a user be converted to an agent or agents in an agent based system while maintaining confidentiality of user information? A user role carries out activities in a system where confidentiality of user information is critical Goals: to achieve optimum performance and maximize gains by taking decisions based on outcome of activities carried out Responsibilities: the responsibilities of this role involve carrying out both non trivial operational tasks and making concluding decisions based on the execution of the tasks carried out. User specific information is used in making the decisions. However, the user information should not be included in the execution of the operational tasks for security and confidentiality reasons This pattern describes an approach for translating a role into agents. It prescribes translating a complex user with sensitive data into two types of agents which are User Agent and Task Agents. The pattern specifies the relationship and control that should exist between these two types of agents Know Uses section mentions the use of the pattern in various scenarios. It includes minimum two examples from different domains The interaction between the assistant agent and the task agents has to be analyzed, modeled and implemented Adaptation/Integration: a user role can be translated into more than one assistant agents depending on the complexity and volume of the user information and decision making process Agent as mediator

A Comparative Study for Ontology and Software Design Patterns

3.4

241

Data Model Patterns

Data Model Patterns help modelers to develop quality models by standardizing common and well-tested solutions for reuse [3]. The objective of data model pattern is to explain a starting point for data modelers [23]. A data model pattern can be implemented by adding additional attribute to any entity in a model or by adding a new entity or a relationship to an existing model. David Hay presented a Universal Data Model in [24]. It is a theoretical model which explains the basic principles of a data model pattern. See Fig. 8:

Fig. 8. Universal data model [19, 23]

In [38], Hay mentioned conventions for building data models. These conventions are guidelines for creating new patterns. They help in establishing a framework which data modelers can follow to reuse data model patterns. The modeling conventions are divided into three levels; Syntactic Conventions, Positional Conventions and Semantic Conventions [19]. Syntactic Conventions: This is the first type of conventions of modeling and deals with the symbols to be used. In the process of syntactic convention evaluation, the crucial point to remember is that there are two audiences in data modeling. The first audience is that community of users which use the models and their descriptions for the verifying whether or not the environment and requirements are actually understood by the analysts. The set of systems designers is the second audience. They make use of the rules of business as implied by the models to be the basis on which their design of computer systems is based [25]. Positional Conventions: This is the second type of Data modeling convention and dictates how entities of a model are laid out. They are concerned with the organization of elements and the overall structure of a model [25]. Semantic Conventions: Semantic conventions are those conventions that address the question of how can the meaning of a model be conveyed. These conventions help to represent common business scenarios in a standard way [25].

242

3.5

Z. Ahmed et al.

Template of Content Ontology Design Patterns

Ontology design patterns are similar to software design patterns. The core idea of describing software design patterns is to use a template and collect them by means of a catalogue. In order to describe ODPs we can use a similar approach as used in software engineering but the difference is that, the template used for the presentation has been optimize for the web and defined in an OWL annotation schema. It is the same used on the semantic web portal http://www.ontologydesignpatterns.org. This part contains a template of Content ODPs which is composed of the following information fields, defined in the annotation schema [2, 5, 26] (Table 7):

Table 7. Content ontology design pattern template [2, 5, 19] Elements Name

Submitted by Also known as Intent Domain Competency question

Solution description Reusable OWL Building Block

Consequences: Scenarios Known uses

Other references

Description It contains the name of the pattern. The names of patterns should be descriptive and unique names that help in identifying and referring to the patterns This part of the template includes author names. In the portal it gives the link to the author page It gives the alternative names for the ODP, since it might be possible that the pattern has some other name but this part is not compulsory This part of the template describes the goal of the ODP. Intent is a description of the goal behind the pattern and the reason for using it This part of the template concerned with the area, domain and where the ODP is applicable It contains a list of competency questions expressed in natural language that are covered by the pattern. A competency question is a classical way of capturing a use case. A competency question is a simple query which an ontology engineer can submit to knowledge base to perform a certain task It describes how the given pattern provides the solution to a design problem in a certain context It is a reusable representation of the pattern. This part is basically the implementation of the design pattern. It contains the URI of the OWL implementation of the content pattern, i.e. the reusable component available for download This part of the template contains a description of the benefits and/or possible trade-offs when using the ODPs Giving examples or scenarios where the given pattern implemented This part of the template gives examples of real ontologies where the ODP is used. This part of template is the example of real usages of the pattern This part of the template contains references to resources (e.g. papers, theories, and blogs) that are related to the knowledge encoded in the ODPs (continued)

A Comparative Study for Ontology and Software Design Patterns

243

Table 7. (continued) Elements Examples

Description This field contains a link of an example owl file which is reusable. The example owl file presents a possible scenario which may sometime also include a UML diagram of classes and their relationships Extracted from Contains the URI (if any) of the ontology from which the pattern has been extracted Reengineered from It contains the name of the reference ontology which has been used reused in the pattern Has components This field refers to components of the Content ODP which are in turn ODPs themselves Specialization of This part of the template refers to ontology elements or ODPs. The specialization relation between ontology elements of ODPs consists of creating subclass of some ODP class and/or sub properties of some ODP properties Related ODP This part contains the names of the patterns which related to the current pattern based on generalization, specialization or composition. It also mentions other patterns that are used in corporation with the current pattern Elements This part of the template describes the elements (classes and properties) included in the ODP, and their role within the ODP Diagram representation This part of the template depicts a graphical representation of the ODP Additional information In Additional information authors provided that informatuion which is not avalible in the rest of the template

Example (See Table 8).

Table 8. Classification pattern [4, 19] Elements Name Submitted by Also known as Intent

Domain Competency question

Description Classification ValentinaPresutti To represent the relations between concepts (roles, task, parameters) and entities (person, events, values), which concepts can be assigned to. To formalize the application (e.g. tagging) of informal knowledge organization systems such as lexica, thesauri, subject directories, folksonomies, etc., where concepts are first-order elements General • What concept is assigned to this entity? • Which category does this entity belong to? (continued)

244

Z. Ahmed et al. Table 8. (continued)

Elements Solution description Reusable OWL Building Block Consequences:

Scenarios Known uses Web references Other references Examples Extracted from Reengineered from Has components Specialization of Related ODP Elements

Description http://www.ontologydesignpatterns.org/cp/owl/classification.owl It is possible to make assertions about e.g., categories, types, roles, which are typically considered at the meta-level of an ontology. Instances of Concept reify such elements, which are therefore put in the ordinary domain of an ontology. It is not possible to parametrize the classification over different dimensions e.g., time, space, etc. Mac OSX 10.5 is classified as an operating system in the Fujitsu-Siemens product catalog

http://www.loa-cnr.it/ontologies/DUL.owl

• Concept (owl:Class). A concept is a Social Object. The classifies relation relates concepts to entities at some time • Entity (owl:Class). Anything: real, possible, or imaginary, which some modeller wants to talk about for some purpose • classifies (owl:ObjectProperty). A relation between a Concept and an Entity, e.g. the Role ‘student’ classifies a Person ‘John’ • is classified by (owl:ObjectProperty). A relation between a Concept and an Entity, e.g. ‘John is considered a typical rude man’; your last concert constitutes the achievement of a lifetime; ‘20-year-old means she’s mature enough’ Diagram representation Fig. 9 Additional information

Fig. 9. Class diagram

A Comparative Study for Ontology and Software Design Patterns

245

4 Study Analysis and Results The work starts with the literature review. Our first task was to study the current patterns that exist in other fields and compare them to ontology design patterns’ templates to analyze the difference. 4.1

Comparison of Pattern Templates

The basic knowledge of problems is in software engineering modeled by conceptual models that are known as Analysis patterns. The patterns could be illustrated using a UML notation [3]. For Example the Analysis Pattern is implemented by using a UML Class diagram, Sequence Diagram and State Diagram which were described in section. Looking at an example of an analysis pattern, one can easily gauge the above relationship between ontologies and Software Patterns, and other patterns used in Computer Science. This similarity is evident even in, for example, graphical representations of an Analysis Pattern and an ontology pattern [3]. 4.2

Comparison of Templates of Software Patterns and Content ODPs

This comparison is between the templates of software patterns and content ontology design patterns. We have described the elements of both software pattern templates and content ODPs in this chapter in Sects. 3.1 to 3.5, which provide us with a good starting point for comparison. There are several types of software patterns that are available and several techniques to present them but we selected Analysis Patterns, Architecture Patterns and Design Patterns. The template of a pattern is a standard way of representing a pattern. In a broad sense, a pattern template has four essential elements. These elements are: Name, Problem, Solution and Consequences as describe in Sect. 3. This comparison is based on similarities and differences between the Content ODP template and software pattern templates. We compare each element of the ODP template with software pattern templates. Comparison is based on the names, content and the overall presentation of the template. The parts described in Table 9 are considered as basic parts of a pattern. The description of these parts is stated. The table shows how these basic parts are described in software patterns and ODP templates. Context: In Software Patterns, the section context describes the situations where the given pattern is applied. Every pattern has a context based on its application area. In content ODPS, context is defined in the form of domains and scenarios where the pattern is applicable. Problem: For Analysis Patterns, Design Patterns and Architectural Patterns, the section Problem or Motivation describes the problem that can be solved by implementing these patterns. In analysis patterns, the section problem describes some generic use cases. In Design Patterns, some of the patterns use scenarios for giving more description of patterns. Such description should be able to clarify the details of the problem. In ODPs, the Problem is described by Competency Questions. Competency Questions consists of a list of competency questions expressed in natural language that

246

Z. Ahmed et al. Table 9. Representing basic elements of other patterns [19]

Pattern type Ontology design pattern

Context Domain

Problem Competency question

Analysis pattern

Context

Problem

Design pattern

Applicability Motivation

Architecture pattern

Context

Problem

Solution Consequence Example OWL Building Block, Consequence Scenario, Example Element, Solution (OWL description and file) Graphical Representation Solution Consequence Example, example resolved Structure, Participant Not provided Sample code Collaboration, implementation and sample code Solution Not provided Not provided

are covered by the pattern. The section Competency Question describes the problem which is to be solved by the ODP. Competency questions are the requirements that an ontology should fulfill. Solution: In Analysis Patterns, Architecture Patterns and Design Patterns, the section Solution gives a fundamental solution principle to be used in the pattern for solving a problem. In Analysis Patterns, the solution is described by using UML diagram and also a brief description of different parts of the pattern is given. In Design Pattern, the sections structure, participants and collaboration describe the solution. The section Structure is a graphical representation of the classes in the pattern which use the notation of the object modeling technique (OMT). It also uses interaction diagrams to illustrate the sequence and collaboration between the objects. The section participants give a detailed description of the classes and objects and their responsibilities. The section collaboration describes how the participants collaborate to carry out their responsibilities. In Architecture pattern, the section solution gives the description, using text and a graphical representation, as to how we can achieve the intended goals and objectives. It also describes the variants and specializations of the solution. It gives the description of people and computing actors and their collaboration. In ODPs, the section solution consists of four parts: OWL Building Block, Elements, Solution description and Graphical Representation. The section Solution description describes how a given pattern can solve the problem in the context. The section OWL Building Block contains an OWL ontology with reusable classes and reusable properties. The section Elements briefly describes Classes and Properties of the pattern implementation and the role of these classes and properties within the ODP. The Graphical representation gives a visual presentation of the pattern classes and their relations. 5Consequence: In both software patterns and content ODPs, the section consequence describes the possible benefits and limitations on the solution after using the pattern. What are the results of using the pattern? In some ODPs it is more about unexpected consequences and limitations.

A Comparative Study for Ontology and Software Design Patterns

247

Example: For Analysis Patterns, Design Patterns and Architecture Patterns, the section Example gives the real world example, which shows the existence of problems and needs for the patterns. For Analysis Patterns and Design Patterns the section example consists of two parts. In the first part the problem and in second part the implementation. In content ODPs, example is given in the form of a scenario and an example OWL file which is the OWL implementation of the scenario. Table 10 describes the common elements of the template of content ontology design patterns and software patterns (analysis patterns, design patterns and architecture patterns). These elements are described in the background chapter with details in Sect. 2.3.2. The section motivation of design pattern is similar to the section problem of other software patterns. In Table 10 the vertical line had different columns that shows the label of different pattern. The horizontal line shows the template heading and symbol X describes the presence of common element.

Table 10. Representing common elements of other patterns [19] Template heading Name Known uses Intent Consequences Also known as Classification

Ontology design pattern     

Analysis pattern    

Design pattern      

Architecture pattern   



The template of content ODPs has developed by following the template of software design patterns. The comparison of both patterns reveals that both patterns have a lot of similarities and the current template of content ODPs includes all the elements of software patterns except ‘forces’. Forces define constraints and problems that can affect the solution. In content ODPs, only the benefits and/or possible trade-offs when using the ODPs on the initial problem are mentioned. Apart from the software patterns, the content ODPs has some additional parts which have been added according to the pattern requirements. Unique Sections: Content ODPs have some unique sections, which are not described in other Software Pattern Templates. These sections are: EXTRACTED FROM, REENGINEERING FROM and HAS COMPONENTS, as mentioned in the background chapter in Sect. 2.5. 4.3

Comparison of Data Model Patterns and Ontology Design Patterns

Data model patterns and ODPs are different to each other because data model patterns are presented only in graphical form while ODPs have much more detailed graphical and textual description. While content ODPs have an official catalogue where users can select from a list of patterns, there is no official catalogue for data model patterns.

248

Z. Ahmed et al.

The template of an ODP has a description of different parts of the pattern which is presented in the graphical and textual form. To implement an ODP, an ontology engineer has to study the description to understand a pattern. A data model pattern is presented in the form of an UML diagram. The diagram is built by following conventions which are a set of rules. These conventions standardize a data model hence makes it easier for a data modeler to reuse a pattern. The reusability of both patterns depends on different factors. Content ODPs can also be used directly as they are, just like data models, although they are usually specialized but this depends on their generality. In data model patterns, it is up to the data modeler to decide whether to use a real model to make some minor adjustments or to use a completely abstract model of a problem. The similarities of data model patterns and ODPs lie in the graphical representation of a problem. Also as we saw in Table 11, different parts of both patterns can be mapped to each other hence knowledge from data model patterns can be translated into ODPs. The use of conventions in data model patterns makes it easier to understand the diagram. These conventions and practices can be translated into guidelines for creating more uniform diagrammatic notations for ODPs. Table 11. Mapping of elements of data model pattern and ontology design pattern [19, 22] Data model pattern Entity Attribute Subtype/supertype Relationships Mutually exclusive sets

Ontology design pattern Concept Relation to “attribute” Subsumption hierarchy Relations Disjoint concepts

Overall, the comparison of software patterns and data model patterns with content ODPs shows that there are a lot of similarities in the templates of software patterns and content ODPs. The difference lies in the presentation of the content. The graphical representation of content ODPs can be made uniform by defining conventions as it is done in data model patterns. In an experiment, the knowledge from data model patterns was translated into ontology design patterns by mapping the parts using the KAON tool [2]. Table 11 shows the mapping between the different parts. The above mapping shows that the knowledge stored in the data model patterns can be translated into ODPs. This knowledge can be reused to create new ODPs from existing data model patterns.

5 Conclusions The purpose of this study was to improve the presentation of content ontology design patterns. This research is part of an effort to solve the larger problem of engineering high quality ontologies. One possible solution to this problem is to introduce reuse in ontology engineering which can standardize the ontology development process and

A Comparative Study for Ontology and Software Design Patterns

249

reduce the time and effort involved in it. ODPs are considered best practice to achieve this objective. To encourage the use of ODPs for ontology development, their presentation must be explicit and precise. Suggestions are based on the comparison of the different patterns with the ontology design patterns. Based on the literature review, we propose following improvements in the current template for content ODP: The Graphical Representation should have a more uniform diagrammatic notation. This can be done by defining standards or conventions which ontology engineers can follow while creating patterns. As guidance, the conventions used in data model patterns can be studied to define similar conventions for content ODPs. Also, namespaces provided in the Graphical Representation section should be made explicit in the Additional Information section. Scenarios should be more explicit. While scenarios presented in the General Description section are simple, they can be more complex in general. Further, the scenario section should describe the binding from concrete elements to the pattern elements. At ODP portal, scenarios of several patterns are missing. They must be included in each pattern because they were considered as important parts. Software patterns include the implementation section with the solution of complex problem. Implementation help user to understand the pattern. This research is part of an effort to solve the larger problem of engineering high quality ontologies. One possible solution to this problem is to introduce reuse in ontology engineering which can standardize the ontology development process and reduce the time and effort involved in it. ODPs are considered best practice to achieve this objective. To encourage the use of ODPs for ontology development, their presentation must be explicit and precise.

6 Future Works This study will be improved through experiment. Determine how well the current ODP template supports the understanding and usage of Content ODPs. To get experts opinions on the current structure of the Ontology design pattern template. There may be certain information that an ontology user need to understand a pattern but it is not available in the description. This research will improve by examine the missing information in the current ontology design pattern templates.

References 1. Uschold, M., Gruninger, M.: Ontologies and semantics for seamless connectivity. ACM SIGMOD Record 33(4), 58–64 (2004) 2. Blomqvist, E.: Fully automatic construction of enterprise ontologies using design patterns: initial method and first experiences. In: OTM Confederated International Conferences On the Move to Meaningful Internet Systems. Springer (2005) 3. Gruber, T.R.: A translation approach to portable ontology specifications. Knowl. Acquis. 5 (2), 199–220 (1993)

250

Z. Ahmed et al.

4. Presutti, V., Gangemi, A.: Content ontology design patterns as practical building blocks for web ontologies. In: International Conference on Conceptual Modeling. Springer (2008) 5. Blomqvist, E.: State of the Art: Patterns in Ontology Engineering (2004) 6. Gangemi, A., Presutti, V.: Ontology design patterns. In: Handbook on Ontologies, pp. 221– 243. Springer (2009) 7. Pree, W.: Meta patterns—a means for capturing the essentials of reusable object-oriented design. In: European Conference on Object-Oriented Programming. Springer (1994) 8. Doran, P.: Ontology reuse via ontology modularisation. In: KnowledgeWeb Ph.D. Symposium. Citeseer (2006) 9. Prabhu, V., Kumara, S., Kamath, M.: Scalable Enterprise Systems: An Introduction to Recent Advances, vol. 3. Springer Science & Business Media, New York (2012) 10. Rebstock, M., Janina, F., Paulheim, H.: Ontologies-Based Business Integration. Springer Science & Business Media, Heidelberg (2008) 11. Guarino, N.: Semantic matching: formal ontological distinctions for information organization, extraction, and integration. In: Information Extraction a Multidisciplinary Approach to an Emerging Information Technology, pp. 139–170. Springer (1997) 12. Fernandez, E.B., Yuan, X.: An analysis pattern for reservation and use of reusable entities. In: Proceedings of PLoP (1999) 13. Buschmann, F., et al.: Pattern-oriented software architecture: a system of patterns (1996). Part II, 2001 14. Kampffmeyer, H., Zschaler, S.: Finding the pattern you need: the design pattern intent ontology. In: International Conference on Model Driven Engineering Languages and Systems. Springer (2007) 15. Cooper, J.W.: The Design Patterns Java Companion, vol. 218. Addison-Wesley, Upper Saddle River (1998) 16. Devedzic, V.: Ontologies: borrowing from software patterns. Intelligence 10(3), 14–24 (1999) 17. Gamma, E.: Design Patterns: Elements of Reusable Object-Oriented Software. Pearson Education India, Delhi (1995) 18. Fernandez, E.B.: Building systems using analysis patterns. In: Proceedings of the Third International Workshop on Software Architecture. ACM (1998) 19. Lodhi, S., Ahmed, Z.: Content Ontology Design Pattern Presentation (2011) 20. Fernandez, E.B., Liu, Y.: The account analysis pattern. In: EuroPLoP (2002) 21. Buschmann, F., et al.: Pattern-Oriented System Architecture: A System of Patterns, pp. 99– 122. Wiley, Chichester (1996) 22. Oluyomi, A.O.: Patterns and protocols for agent-oriented software development (2006) 23. Silverston, L., Inmon, W.H., Graziano, K.: The Data Model Resource Book: A Library of Logical Data Models and Data Warehouse Designs. Wiley, Hoboken (1997) 24. Hay, D.: Data Model Patterns Conventions of Thought. Dorset House, New York (1996) 25. Hay, D.C.: A comparison of data modeling techniques (1999) 26. Gangemi, A.: Ontology design patterns for semantic web content. In: International Semantic Web Conference. Springer (2005)

Biomedical Applications

MainIndex Sorting Algorithm Adeel Ahmed1, Muhammad Arif1(&), Abdul Rasheed Rizwan2, Muhammad Jabbar1, Zaheer Ahmed1, and Muhammad Sami Ullah1 1

Department of Computer Science, University of Gujrat, Gujrat City, Gujrat, Pakistan {m.arif,adeel.ahmed,m.jabbar, zaheer,msamiullah}@uog.edu.pk 2 Department of Computer Science, Virtual University of Pakistan, Lahore, Pakistan [email protected]

Abstract. Sorting algorithm remained hot topic in computer science from the birth of computer science to achieve maximum performance. Fortunately this achievement became possible due to good and fast sorting algorithms, such as heap sort, merge sort, radix sort and other sorting algorithms. Till this achievement is also under research to find more efficient algorithms. In sorting algorithm arrays and link list data structures are commonly used. We know arrays are efficient if we need consecutive kind of data structure and link lists are useful when we need to add and remove items in the data structure. In other word we can say both data structures have own its merits and demerits. So in our sorting algorithm we are going to use both kinds of data structure. We will use in our MainIndex sorting algorithm arrays as the MainIndex and link list as sorting cells. MainIdex sorting algorithm need some kind of information just the length of the number which is going to sort and the value of the number which is going to sort in sorting cell. Keywords: MainIndex
Sorting cell
Selection sort
Link list
Merge sort
Bubble sort
Quick sort

1 Introduction A sorting algorithm exchanges the numbers in a specific order according to the fit needs of the users. Any sorting algorithm has merging and searching functions. Some other exist which used the functions to sort the numbers [1]. The data is mostly taken from an array which is faster in sense of random access [2]. Link list provide a fast method when we need to add or delete node for numbers against arrays [3]. Sorting algorithms remain hot topic since the birth of computer due to time complexity. Time complexity is the total time required by an algorithm of the program to complete its given task [4]. The time complexity is measurement by big O notation of the algorithm. The numbers of elements are measured to compute the complexity time. Sorting algorithms are the fundamental concepts for a computer programmer, computer scientist and IT

© Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_20

254

A. Ahmed et al.

professionals [5]. When we search in database mostly used ORDER BY clause, using this query actually we are using the sorting algorithm. A common example is our phone book of address, we save our contact numbers in the mobile in a different order, but when we need to find a contact they are listed into a sequence [6]. First look we should compare various sorting algorithms in dealing with large set of data, in large set of data some algorithms become very slow in practical use. Next look is about the memory usage of the algorithms, mostly faster algorithms takes large memory rather than slow one algorithm [7]. If we look about the progress in the memory usage of today, large memory is not main issue, but issue is faster algorithms [8]. But we even imagine which algorithm is fastest; the speed of the algorithm depends on the working and environment where the sorting is done. In sense which kind of data is being to sort? We cannot use the same algorithm for sorting 200 numbers on such kind of database which cannot fit into memory, so such kinds of algorithms are designed in different ways.

2 Related Works Today many programming languages are providing built in functionalities which are usable for sorting numbers [9]. Java API, .Net framework and C++ all provides built-in sorting techniques [10]. For different kinds of data types such as floating numbers, integer number and characters need to sort according to the requirements. But sometime we need a generic kind of sorting algorithms for generic data structure [11]. This thing leads us to a complicated problem. But OOP provides us to solve such kinds of problems [12]. So many good algorithms are designed but no one algorism is Silver Bullet. At this point we see the complexity of five algorithms. As given in the Table 1.

Table 1. Complexity of the algorithms Sorting algorithm Bubble sort Insertion sort Merge sort Heap sort Quick Sort

2.1

Worst case O(n2) O(n2) O(n log n) O(n log n) O(n2)

Average case O(n2) O(n2) O(n log n) O(n log n) O(n log n)

Best case O(n) O(n) O(n log n) O(n log n) O(n log n)

Bubble Sort

The bubble sort algorithm is a well-known sorting algorithm. Its belongs to O (n2) sorting algorithm. But bubble sort is not very efficient for sorting of large amount of data, other hand it is stable and adaptive algorithm [13].

MainIndex Sorting Algorithm

255

Bubble Sort Algorithm for i = 1 to n; swapping = false for j = n: i+1; if a [j] < a [j-1]; swap a[j,j-1]; swapping = true; if not swapping; break; end 2.2

Insertion Sort

Insertion sort is elementary sorting algorithm and its time complexity in wore case is O(n2). Insertion Sort is a good choice when data is nearly related to each other due to adaptive algorithm. But it is not good when data size is small. Insertion Sort is recursive by nature when the problem size is small [14].

Insertion Sort Algorithm for i = 2 to n; for (k = i; j > 1 and a[j] < a[j-1]; j--); swap a[j,j-1]; end 2.3

Merge Sort

Merge sort bases on divide and conquer algorithm. It divides the problem into two parts and sorts them, after sorting them it merges them. Merge Sort is best choice for different kinds of environments for example when sorting with link list, stability and fast result is required. But it is expensive when the problem is in array [15]. 2.4

Heap Sort

Heap Sort algorithm bases on binary heap. In some sense it is similar to selection sort where we find the maximum value in start and rearrange it at the end; we repeat the same process for remaining numbers in the problem [16]. 2.5

Quick Sort

Quick Sort also works in Divide and Conquer method just as Merge sort. It takes a number as pivot and make the partition according to the selected pivot. Selecting a pivot is grand work in this method. Different kinds of ways are to select the pivot. • Pick first number as pivot. • Pick last number as pivot.

256

A. Ahmed et al.

• Pick random number as pivot. • Pick middle number as pivot [17].

3 Problem Statement We took an overview on different kinds of sorting algorithms and saw each algorithm is using a specific data structure such as arrays or link list. So every sorting algorithm has own its pros and cons, some need more memory but they are sharp in sorting the numbers, some algorithm needs less memory but it takes more time to sort the numbers. Different sorting algorithms take different time to sort the numbers. Every algorithm behaves different according to the nature of the problem. We are going to discuss specific method which behave constant in every nature of sorting data and takes less time against all sorting algorithms. In Main Index sorting algorithm we will use both kinds of data structure such as link list and arrays according to their efficient point of view.

4 The Proposed Solution Now we are going to discuss a new sorting algorithm, we have following unsorted array. 32, 18, 45, 56, 22, 64, 8, 88, 67, 105, 77, 440, 11, 90, 16 We have these unsorted serials of numbers in array. To sort them in order first of all we should compute the maximum length in unsorted numbers. We are not going to find the maximum number in unsorted numbers. In our unsorted list 32 is the first number, in this we need two kind of information number one the value of first digit which is 3 in case of 32. Second we need the length of 32 which is 2. The length 2 numeric numbers should store in LIndex2. Later we will see what mean of LIndexN. Before going in detail of MainIndex first we should take an over view of the sketch of the MainIndex. Following is the basic detail figure of the MainIndex. As shown in the Fig. 1.

Fig. 1. Sketch of the MainIndex

MainIndex Sorting Algorithm

4.1

257

MainIndex

MainIndex has a array from 1 to 9, this section will use appropriate position according to the value of the number, let’s take 32 has example here again, 32 has first digit 3 so it should enter in array of 3 in MainIndex. The arrow is a pointer which is pointing to LIndex1. LIndex1 mean length of index 1 which is core structure with the MainIndex. All those numeric numbers which have its length one those should store in this index. Other LIndexN should create when a numeric number has different length against already created LIndexN. But LIndex1 should create automatically when MainIndex starts. 4.2

LIndexN

L mean length of the numeric number which is going to store in the sorting sequence, N is index number. Every LIndexN has its own sorting cells which should program in link list; we need compression of existing numbers with the new coming number, after compression it should store in its appropriate place. Internal structure of sorting cells is as following. It is given in the Fig. 2.

B) then a signal with a single component is more likely to have a component only. Otherwise, the analyzed signal has more components. This statement is verified by the data available in Table 2. C. Information Measures Table 3 presents the numerical values of the absolute (RH) and differential Renyi entropies (DRH). The differential entropy is obtained by subtracting the entropy of the case #1 from the absolute values of all others entropies. Two distributions are considered, namely WVD and PWVD. Figure 7 shows a comparison between entropies, as solution describing the complexity of the analyzed signal, in all ten considered cases. It is important to notice the superiority of the entropies. If the complexity of the signal is growing (at least as the number of components) then the entropy is rising, too. Moreover, differential entropies

Table 3. Renyi entropies No WVD RH 1. −0.2012 2. −0.2075 3. −0.2075 4. −0.2075 5. +1.1465 6. +1.1469 7. +1.1469 8. +2.0483 9. +2.0895 10. +2.5921

DRH 0.0 −0.0063 −0.0063 −0.0063 +1.3477 +1.3480 +1.3481 +2.2495 +2.2907 +2.7933

PWVD RH −0.0548 −0.1896 −0.2030 −0.2055 1.2647 0.9987 0.9262 2.0225 1.7881 2.4355

DRH 0.0 −0.1349 −0.1482 −0.1507 +1.3195 +1.0535 +0.9810 +2.0772 +1.8429 +2.4903

442

D. Aiordachioaie

Fig. 7. Evolution of the main descriptors (entropies), information-based

are able to predict the number of the components in the analyzed signal, as it was described in Sect. 3B. The information obtained from entropies combined with the ratios of statistical parameters, as described above, could drastically improve the results of the signal analysis, both on qualitative (as number of components) and quantitative aspects (localization and width in time and frequency). D. Detection and Filtering of Interferences With reference to the system of Fig. 1, let TInxn be a block image for testing. Let bInxn (bad image) be the block image with artifacts and let NInxn (normal image) be the block image with no interferences. The masking decision is based on the decision rule: IF corrðTI; bIÞ ¼ max AND THEN FILTER TI

corrðTI; NIÞ ¼ min

ð24Þ

where corr(.,.) is the statistic correlation operator working on square matrices. Figure 8 presents four block images: two for clean (correct) images, x1 and x4, and two block images with interferences, x12 and x34. The interference images have two specific features, which could be used in automatic detection and classification: (i) it contains a lot of parallel ellipses; (ii) The main diagonals are parallel to the segment which connects the centers of the closest neighbored components. This set of features could be used to distinguish the real from false components in time-frequency images. In this case, the self-checking procedure could be used before calling a classifier for recognition purposes. Figure 9 presents the TFI for case # 8 and the results of the cross-correlation between the TFI and the bad block assigned to x12. The maximum cross-correlation indicates the position of the most similar block, which should be processed. Finally, Fig. 10 shows the effect of removing two bad blocks (block with interferences) from the raw time-frequency image (TFI), for the same case (#8).

On Time-Frequency Image Processing for Change Detection Purposes

443

Fig. 8. Samples from the feature image data base: with and free of interference

Fig. 9. Raw TFI image and the cross-correlation image between x12 and raw TFI

The complete set of experiments can fix the performance of the system. The results are as expected for all cases with double components. For the cases with more components, the results depend on the distance between components. If some components are in the middle, between two or more components, then they are modified and the reconstruction becomes quite difficult with this system.

444

D. Aiordachioaie

Fig. 10. Raw and filtered TFI images, case 8

5 Conclusions A system meant to analyze, detect and filter artifacts in images coming from time-frequency transforms was considered and discussed, in the context of change detection and diagnosis of industrial equipment, which generates mechanical vibrations. The complexity of the input signal is well described by the Renyi entropy, which increases when the complexity of the analyzed signal is growing (number of components, mainly). The average parameters of time and frequency are good descriptors only for mono-components signals. Otherwise, they refer to the average content. The detection system uses the correlation between small (blocks) images and scanned input images. The raw time-frequency image is decomposed in basic (small) blocks of size nxn = 100  100. The decomposition could be static, by scanning the raw image on vertical and horizontal axis, in steps of one or several pixels, or could be dynamic, by changing the size of the scanning steps, based on the content of the image. Sparse blocks will generate higher scanning steps. The dynamic scanning speeds up the analysis of the image. The interference images have two specific features, which could be used in automatic detection and classification: (i) it contains a lot of parallel ellipses; (ii) The main diagonals are parallel to the segment which connects the centers of the closest neighbored components. This set of features could be used to distinguish the real from false components in time-frequency images. The system needs a training stage (time) to complete a database with what means good and bad features in the analyzed images. This process is supervised and could be run all the time. The similarity checking between the content of the raw image and the selected features for database is made on cross-correlation basis. The performance of the system is acceptable for signals with well-separated components, in time and frequency, which is the case of many real faults in the industry of rotating machines. If the components and the artifacts interfere, the system cannot work as expected, and generates errors. It could be involved in estimating the values of the real components, but this aspect is not considered here.

On Time-Frequency Image Processing for Change Detection Purposes

445

Acknowledgement. The authors thank the Executive Agency for Higher Education, Research, Development and Innovation Funding (UEFISCDI) for its support under Contract PN-II-PTPCCA-2013-4-0044 (VIBROCHANGE).

References 1. Isermann, R.: Supervision, fault-detection and fault-diagnosis methods. CEP 5(5), 639–652 (1997) 2. Patton, R.J., Frank, P., Clark, R. (eds.): Fault Diagnosis in Dynamic Systems – Theory and Application. Prentice Hall, Upper Saddle River (1989) 3. Basseville, M., Nikiforov, I.: Detection of Abrupt Changes – Theory and Application. Prentice Hall, Upper Saddle River (1993) 4. Gustafson, F.: Statistical signal processing approaches to fault detection. ARC 31, 41–54 (2007) 5. Basseville, M.: On-board component fault detection and isolation using the statistical local approach. Publication Int. No. 1122, IRISA (1997) 6. Timusk, M., Lipsett, M., Mechefske, C.K.: Fault detection using transient machine signals. Mech. Syst. SP 22, 1724–1749 (2008) 7. Venkatasubramanian, V., Rengaswamy, R., Yin, K., Kavuri, S.N.: A review of process fault detection and diagnosis, part I: quantitative model-based methods. Comput. Chem. Eng. 27, 293–311 (2003) 8. Wang, W.J., McFadden, P.D.: Time-Frequency Domain Analysis of Vibration Signals for Machinery Diagnostics, (III) The Present Power Spectral Density, University of Oxford, Report No. OUEL 1911/92 (1992) 9. McFadden, P.D., Wang, W.: Time-Frequency Domain Analysis of Vibration Signals for Machinery Diagnostics. (I) Introduction to the Wigner-Ville Distribution, University of Oxford, Report No. OUEL 1859/92 (1990) 10. The VIBROCHANGE project, Experimental Model for Change Det. and Diag. of Vibrational Proc. Using Advanced Measuring and Analysis Techn. Model-Based (2015). www.etc.ugal. ro/VIBROCHANGE 11. Cohen, L.: Time-frequency distributions - a review. Proc. IEEE 77(7), 941–980 (1989) 12. Hlawatsch, F., Boudreaux-Bartels, F.: linear and quadratic time-frequency signal representations. IEEE Sig. Process. Mag. 9, 21–67 (1992) 13. Flandrin, A.P., Gonçalvès, P., Lemoine, O.: Time-Frequency Toolbox For Use with MATLAB, CNRS (France)/Rice Univ. (USA) (1995–1996) 14. Auger, F.: Representations Temps-Frequence Des Signaux Nonstationnaires: Synthese Et Contributions. Ph.D. thesis, Ecole Centrale de Nantes, France (1991) 15. Flandrin, P.: Temps-frequence, Hermes, Trait des Nouvelles Technologies, serie Traitement du Signal (1993) 16. Popescu, T., Aiordachioaie, D.: Signal segmentation in time-frequency plane using renyi entropy - application in seismic signal processing. In: 2nd International Conference on Control and Fault-Tolerant Systems, SysTol-2013, 9–11 October, Nice, France, pp. 312–317 (2013) 17. Aiordachioaie, D.: Signal segmentation based on direct use of statistical moments and renyi entropy. In: 10th International Conference on Electronics, Computer and Computation (ICECCO 2013), Istanbul, Turkye, pp. 359–362 (2013) 18. Barbarossa, S.: Analysis of multicomponent LFM signals by a combined wigner-hough transform. IEEE Trans. Sig. Process. 43(6), 1511–1515 (1995)

Adaptive Multi-round Smoothing Based on the Savitzky-Golay Filter J´ ozsef Dombi1(B) and Adrienn Dineva2,3 1

3

Institute of Informatics, University of Szeged, ´ ad t´er 2, Szeged, Hungary 6720 Szeged Arp´ [email protected] 2 Doctoral School of Applied Informatics and Applied Mathematics, ´ Obuda University, 1034 Budapest B´ecsi u ´t 96/b, Budapest, Hungary [email protected] Doctoral School of Computer Science, Department of Information Technologies, Universit´ a degli Studi di Milano, Crema Campus, 65 Via Bramante, 26013 Crema (CR), Italy

Abstract. Noise cancellation is the primary issue of the theory and practice of signal processing. The Savitzky-Golay (SG) smoothing and differentiation filter is a well studied simple and efficient technique for noise eliminating problems. In spite of all, only few book on signal processing contain this method. The performance of the classical SGfilter depends on the appropriate setting of the windowlength and the polynomial degree. Thus, the main limitations of the performance of this filter are the most conspicious in processing of signals with high rate of change. In order to evade these deficiencies in this paper we present a new adaptive design to smooth signals based on the Savitzky-Golay algorithm. The here provided method ensures high precision noise removal by iterative multi-round smoothing. The signal approximated by linear regression lines and corrections are made in each step. Also, in each round the parameters are dynamically change due to the results of the previous smoothing. The applicability of this strategy has been validated by simulation results.

Keywords: Savitzky-Golay filter Iterative smoothing · Denoising

1

·

Adaptive multi-round smoothing

·

Introduction

Many areas of signal processing require highly efficient processing methods in order to achieve the desired precision of the result. In a particular class of tasks, for e.g. chemical spectroscopy, smoothing and differentiation is very significant. An ample number of studies have been revealed the details of the smoothing filters. In 1964 a great effort has been devoted to the study of Savitzky and Golay, in which they introduced a particular type of low-pass filter, the so-called digital smoothing polynomial filter (DISPO) or Savitzky-Golay (SG) filter [17]. c Springer International Publishing AG 2018  V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8 38

Adaptive Multi-round Smoothing Based on the Savitzky-Golay Filter

447

The main advantage of the SG-filter in contrast to the classical filters - that require the characterzation and model of the noise process-, is that both the smoothed signal and the derivatives can be obtained by a simple calculation. Critical reviews and modifications of the original method can be read, for instance in [15,21]. The core of this algorithm is fitting a low degree polynomial in least squares sense on the samples within a sliding window. The new smoothed value of the centerpoint obtained from convolution. There is a rapidly growing literature discussing the properties and improvements of SG filters [1,3,6,7,12,16,24,25]. Also the importance and applicability of a digital smoothing polynomial filter in chemometric algorithms are well established [8,11,22]. The frequency domain properties of SG-filters are addressed in [2,10,18,19]. In [14], the properties of the SG digital differentiator filters and also the issue of the choice of filter length are discussed. Paper [13] concerns the calculation of the filter coefficients for even-numbered data. Also the fractional-order SG differentiators have been investigated, for e.g., by using the Riemann-Lioueville fractional order definition in the SG-filter. For instance, the fractional order derivative can be calculated of corrupted signals as published in [4]. There are several sources and types of noise that may distort the signal, for e.g., eletronic noise, electromagnetic and electrostatic noise, etc. [23]. In the theory of signal processing it is commonly assumed that the noise is an additive white Gaussian noise (AWGN) process. However, in engineering practice often nonstationary, impulsive type disturbances, etc., can degrade the performance of the processing system. Since, for the noise removal issue of signals with a large spectral dynamic or with a high rate of change, the classical SG filtering is an unefficient method. Additionally, the performance depends on the appropriate selection of the polynomial order and the window length. The arbitrary selection of these parameters is difficulty for the users. Usually the Savitzky-Golay filters perform well by using a low order polynomial with long window length or low degree with short window. This latter case needs the repetition of the smoothing. It has also been declared that the performance decreases by applying low order polynomial on higher frequencies. Nonetheless, it is possible to further improve the efficiency. With this goal, in this work we introduce an adaptive smoothing approach based on the SG filtering technique that ensures acceptable performance independent of the type of noise process.

2

Mathematical Background of the Savitzky-Golay Filtering Technique

In this section we briefly outline the premise behind the SavitzkyGolay filtering according to [9]. Let us consider equally spaced input data of n{xj ; yj }, j = 1, . . . , n. The smoothed values derives from convolution, given by gi =

m  i=−m

ci yk+i ,

(1)

448

J. Dombi and A. Dineva

where the window length M = 2m + 1, i = −m, . . . , λ, . . . , m, and λ denotes the index of the centerpoint. The k th order polynomial P can be written as P = a0 + a1 (x − xλ ) + a2 (x − xλ )2 + . . . + ak (x − xλ )k

(2)

The aim is to calculate the coefficients of Eq. (2) by minimizing the fitting error in the least squares sense. The Jacobian matrix is as follows J=

∂P ∂a

(3)

The polynomial at x = xλ is a0 , hence in order to evaluate the polynomial in the window we have to solve a system of M equations which can be written in matrix form J ·a=y (4) ⎞ ⎞ ⎛a ⎞ ⎛y 0 λ−m 1 (xλ−m − xλ ) · · · (xλ−m − xλ )k .. ⎟ ⎜ .. ⎟ ⎜ .. ⎟ ⎜ .. .. .. ⎜ ⎟ . ⎟ ⎜. ⎟ ⎜ . . . ⎟ ⎜ . ⎟ ⎜ ⎟ ⎜ ⎜ ⎜ ⎟ . . ⎜1 ⎟×⎜ . ⎟=⎜ . ⎟ 0 ··· 0 ⎜ ⎟ ⎜ . ⎟ ⎜ . ⎟ ⎜. ⎟ ⎜ . ⎟ ⎜ . ⎟ .. .. .. ⎝ .. ⎠ ⎝ . ⎠ ⎝ . ⎟ . . . . . ⎠ k 1 (xλ+m − xλ ) · · · (xλ+m − xλ ) ak yλ+m ⎛

The coefficients are found from the normal equation in the following writing

so Since

J T (Ja) = (J T J)a

(5)

a = (J T J)−1 (J T y).

(6)

P (xλ ) = a0 = (J T J)−1 (J T y),

(7)

by replacing y with a unit vector in Eq. (6) the c0 coefficient can be calculated as cj =

k+1 

|(J T J)−1 |0i Jij .

(8)

i=1

With a size of (2m + 1)x(k + 1) the G matrix of the convolution coefficients G = J(J T J) = [g0 , g1 , . . . , gj ].

(9)

Figure 1 demonstrates the performance of the classical SG-filter. It can be observed that the smoothing is not precise. To overcome this problem, the following section will present an adaptive strategy (Table 1).

Adaptive Multi-round Smoothing Based on the Savitzky-Golay Filter

449

Fig. 1. Performance of the SG filter. Upper chart: signal with contaminating noise. Lower chart: dotted line - original signal, solid line - smoothed signal, k = 3, M = 35 Table 1. Some SG coefficients. M = 2m + 1 is the window length and k denotes the polynomial degree Savitzky-Golay coefficients M k Coefficients 2*9 2 -0.0909 0.0606 0.1688 0.2338 0.2554 0.2338 0.1688 4 0.0350 -0.1282 0.0699 0.3147 0.4172 0.3147 0.0699 2*11 3 -0.0839 0.0210 0.1026 0.1608 0.1958 0.2075 0.1958 0.1608 5 0.0420 -0.1049 -0.0233 0.1399 0.2797 0.3333 0.2797 0.1399

3 3.1

0.0606 -0.0909 -0.1282 0.0350 0.1026 0.0210 -0.0839 -0.0233 -0.1049 0.0420

Adaptive Multi-round Smoothing Based-On the SG Filtering Technique Multi-round Smoothing and Correction

This new adaptive strategy aims setting automatically the suitable polynomial order and window length at the different frequency components of the signal. Hence, it is possible to avoid the undershoots and preserve the peaks that could be important from different data analysis aspects. Since we perform in the time domain, this method provides efficient results independent of the type of contaminating noise. At first, the classical Savitzky-Golay filtering is performed. Assuming that only the corrupted signal is available, this step serves for revealing the peaks, hence the window length and degree of the polynomial may be arbitrary. After the first smoothing, the coordinates of the local minimum and maximum points can be obtained. From now on, we can also define the d distace vector which contains the number of samples between two neighboring points

450

J. Dombi and A. Dineva

of local minima and maxima. Then, the next step is the separation of the highand low frequency components using the bordering points and setting the proper parameters for the smoothing. The window should match the scale of the signal and the polynomial degree should vary by depending on the framesize and frequency. Since the next fuzzy relation can be defined between the section length; F (dmax >> d¯R ) =

1 1+

e−(δmax −d¯R )

∈ [0, 1]

(10)

where δ¯R stands for the average length of the sections in the current R parts of the signal, while δmax = max(d) in the observed signal. If g(dmax , d¯R ) = 1, the current part of the signal contains high freqency componenst. Hence, the following rules are applied: if 1 > g(dmax , d¯R ) > 0.9 then k = 5, M = nint(0.3δ¯R ) if 0.89 > g(dmax , d¯R ) > 0.75 then k = 4, M = nint(0.5δ¯R ) if 0.75 > g(dmax , d¯R ) > 0.45 then k = 3, M = nint(δ¯R ) if 0.44 > g(dmax , d¯R ) > 0.2 then k = 2, M = nint(0.5Rn )

(11)

else k = 1, M = nint(0.8Rn ), where Rn is the total number of samples of the R part, and we can assign the k and M values to each R part of the signal. The values for the bounds have been determined according to the forlmula 2k modified by experimental results. 3.2

Approximation Using Modified Shephard Method for Corretion

The correction carried out with taking the linear approximation of the obtained signal. Then, it is extracted from the smoothed one. This step reveals the higher deviation, thus the next smoothing procedure can be modified accordint to its result. As we have the coordinates of the local minimum and maximum points and the vector d, we can easily fit a regression line on the points between two local extrema. In this case, the ending and starting points of two consecutive lines do not necessary follow so as to match at the same value. In order to ensure the continuous joining of the lines we can perform this step by appling the Lagrange-multiplicator method given by   (m1 x(i) + b1 − y (i) )2 + (m2 x(i) + b2 − y (i) )2 ⇒ min, (12) xi ∈[x1 ,x2 ]

xi ∈[x1 ,x2 ]

with the following constraints: m1 x2 + b1 − m2 x2 + b2 .

(13)

However, in some cases the peaks can contain the information of interest. Therefore, the form of the peak or valley should be processed with special care. To addres this issue, a modified Shepard - method can be applied. Let us consider

Adaptive Multi-round Smoothing Based on the Savitzky-Golay Filter

451

Fig. 2. Illustration of the problem of joining of the regression lines

the points around the local extrema in radius r. The new values are calculated by weighting according to the neighboring points distance. There are several variations of the Shepard method [20], now let us consider the GM S (Groundwater M odeling System) form below i 2 ( d−d ddi ) . wi = n i 2 ( d−d ) ddi

(14)

i=1

Equation 14 can be trasformed into i 2 ( 1−u ui ) , wi = n u−u ( uj j )2

(15)

j=1

i

(x) in which ui (x) = dd(x) . Now, using the similarity between the form of Eq. 15 and the Dombi operator [5] we can define the following new parametric weighting function: 1 wi = (16) 1−uj λ ui 2 1 + ( 1−ui ) ( uj ) j=1

in which the setting of λ and radius r (where it performs) have the effect on the smoothness of the result (Fig. 2).

4

Simulation Results

The performance of the proposed method have been tested on a noisy signal (see, Fig. 3). The simulation has been carried out by using Matlab 8. Figure 4 shows the approximated signal after the first round. In Fig. 5 the resulted and the original signal can be seen after two rounds. It can be observed that the applied technique can efficiently recover the signal.

452

J. Dombi and A. Dineva

Fig. 3. The corrupted signal

Fig. 4. The approximation of the signal after the first round.

Fig. 5. The recovered (blue) and the original (magenta) signal.

Adaptive Multi-round Smoothing Based on the Savitzky-Golay Filter

5

453

Conclusions

In this paper a new adaptive smooting strategy has been introduced based on the Savitzky-Golay filtering technique. The proposed method allows to evade the main difficulties of the original SG filter by automatically setting the smoothing parameters. Furthermore, for the precise reconstruction of the signal a multi round correction has been applied using the linear approximation of the signal. For the reconstruction of the peaks and valleys that may contain the important information, a new weighting function has been introduced with the combination of the GM S method and the Dombi operator. The applicability and efficiency of this new strategy have been validated by simulation results. Acknowledgement. This work has been sponsored by the Hungarian National Scientific Research Fund (OTKA 105846). The authors also thankfully acknowledge the support of the Doctoral School of Applied Informatics and Applied Mathematics of Obuda University.

References 1. Ahnert, K., Abel, M.: Numerical differentiation of experimental data local versus global methods. Comput. Phys. Commun. 177, 764–774 (2007) 2. Bromba, M.U.A., Ziegler, H.: Application hints for Savitzky-Golay digital smoothing filters. Anal. Chem. 53(11), 1583–1586 (1981) 3. Browne, M., Mayer, N., Cutmore, T.: A multiscale filter for adaptive smoothing. Digit. Sig. Proc. 17, 69–75 (2007) 4. Chen, D., et al.: Digital fractional order Savitzky-Golay differentiator. IEEE Trans. Circ. Syst. 58(11), 758–762 (2011) 5. Dombi, J.: Towards a general class of operators for fuzzy systems. IEEE Trans. Fuzzy Syst. 16(2), 477–484 (2008) 6. Edwards, T., Willson, P.: Digital least squares smoothing of spectra. Appl. Spectrosc. 28, 541–545 (1974) 7. Engel, J., et al.: Breaking with trend in pre-processing? Trends Anal. Chem. 2013, 96–106 (2013) 8. Finta, C., Teppola, P.J., Juuti, M., Toivanen, P.: Feasibility of parallel algorithms and graphics processing unit acceleration in chemical imaging. Chemometr. Intell. Lab. Syst. 127, 132–138 (2013) 9. Flannery, B., Press, H., Teukolsky, A., Vetterling, W.: Numerical Recipes in C: The Art of Scientific Computing, 2nd edn. Cambridge University Press, New York (1992) 10. Hamming, R.W.: Digital Filters, 3rd edn. Prentice-Hall, Englewood Cliffs (1989). Chap. 3 11. Komsta, L.: A comparative study on several algorithms for denoising of thin layer densitograms. Anal. Chim. Acta 641, 52–58 (2009) 12. Krishnan, R., Seelamantula, C.: On the selection of optimum Savitzky-Golay filters. IEEE Trans. Sig. Process. 61(2), 380–391 (2013) 13. Luo, J., Ying, K., Bai, J.: Savitzky-Golay smoothing and differentiation filter for even number data. Sig. Process. 85(7), 1429–1434 (2005)

454

J. Dombi and A. Dineva

14. Luo, J., Ying, K., He, P., Bai, J.: Properties of Savitzky-Golay digital differentiators. Digit. Sig. Proc. 15(2), 122–136 (2005) 15. Madden, H.: Comments of Savitzky-Golay convolution method for least squares fit smoothing and differentiation of digital data. Anal. Chem. 50, 1383–1386 (1978) 16. Persson, P.O., Strang, G.: Smoothing by Savitzky-Golay and legendre filters. In: Rosenthal, J., Gilliam, D.S. (eds.) Mathematical Systems Theory in Biology, Communications, Computation, and Finance, vol. 134, pp. 301–315. Springer, New York (2003) 17. Savitzky, A., Golay, M.J.: Smoothing and differentiation of data by simplified least squares procedures. Anal. Chem. 36(8), 1627–1639 (1964) 18. Schafer, R.W.: On the frequency-domain properties of Savitzky-Golay filter. In: Proceedings of the the 2011 DSP/SPE Workshop, Sedona, AZ, pp. 54–59 (2011) 19. Schafer, R.W.: What is a Savitzky-Golay filter? (lecture notes). IEEE Sig. Process. Mag. 28(4), 111–117 (2011). doi:10.1109/MSP.2011.941097 20. Shepard, D.: A two-dimensional interpolation function for irregularly-spaced data. In: Proceedings of the 1968 ACM National Conference, pp. 517–524 (1968) 21. Steiner, J., Termonia, Y., Deltour, J.: Comments on smoothing and differentiation of data by simplified least squares procedura. Anal. Chem. 4, 1906–1909 (1972) 22. Tong, P., Du, Y., Zheng, K., Wu, T., Wang, J.: Improvement of nir model by fractional order Savitzky-Golay derivation (FOSGD) coupled with wavelength selection. Chemometr. Intell. Lab. Syst. 143, 40–48 (2015) 23. Vaseghi, S.V.: Advanced Digital Signal Processing and Noise Reduction, 4th edn. Wiley, Hoboken (2008) 24. Wayt, H.J., Khan, T.R.: Integrated Savitzky-Gola filter from inverse Taylor series approach. In: Proceedings of the 2007 15th International Conference on Digital Signal Processing (DSP 2007), pp. 375–378. IEEE (2007) 25. Zhao, A., Tang, X., Zhang, Z.: The parameters optimization selection of SavitzkyGolay filter and its application in smoothing pretreatment for FTIR spectra. In: Proceedings of the IEEE 9th Conference on Industrial Electronics and Applications (ICIEA), pp. 516–521 (2014)

Point Cloud Processing with the Combination of Fuzzy Information Measure and Wavelets Adrienn Dineva1,2(B) , Annam´ aria R. V´ arkonyi-K´ oczy3 , Vincenzo Piuri4 , 5 and J´ ozsef K. Tar 1

Doctoral School of Applied Informatics and Applied Mathematics, ´ Obuda University, B´ecsi u ´t 96/b, Budapest 1034, Hungary [email protected] 2 Doctoral School of Computer Science, Department of Information Technologies, Universit´ a degli Studi di Milano, Crema Campus, 65 Via Bramante, 26013 Crema (CR), Italy 3 Department of Mathematics and Informatics, J. Selye University, Komarno, Slovakia [email protected] 4 Department of Information Technologies, Universit´ a degli Studi di Milano, Crema, Italy [email protected] 5 ´ Antal Bejczy Center for Intelligent Robotics (ABC iRob), Obuda University, Budapest, Hungary [email protected]

Abstract. Processing of remotely sensed point clouds is a crucial issue for many applications, including robots operating autonomously in real world environments, etc. The pre-processing is almost an important step which includes operations such as remove of systematic errors, filtering, feature detection and extraction. In this paper we introduce a new preprocessing strategy with the combination of fuzzy information measure and wavelets that allows precise feature extraction, denoising and additionally effective compression of the point cloud. The suitable setting of the applied wavelet and parameters have a great impact on the result and depends on the aim of preprocessing that usually is a challenge for the users. In order to address this problem the fuzzy information measure has been proposed that supports the adaptive setting of the parameters. Additionally a fuzzy performance measure has been introduced, that can reduce the complexity of the procedure. Simulation results validate the efficiency and applicability of this method. Keywords: Point cloud processing · Fuzzy information measure · Signal processing · Denoising

1

Introduction

Recently, there is a growing interest in several areas of engineering practice for the automatic processing methods of three-dimensional remotely sensed models. c Springer International Publishing AG 2018  V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8 39

456

A. Dineva et al.

Most of the 3D scanning devices produce a set of data points of the environment usually defined with their Descartes (x, y, z) coordinates. The resulted data set can contain faulty sensor values of the environment, disturbing noises, etc.,. Since the sampling of the raw point cloud is insufficient for most applications the main steps of processing includes the outlier (extreme values) removal, noise cancellation and smoothig, surface reconstruction and feature extraction. Additionally, depending on the specific requirements of the application further post processing techniques can be used [10]. Classical approaches, for instance linear filtering, can remove the noise, but with weak feature localization ability. To overcome this deficieny nonlinear filters have been proposed [8,16]. Among the classical signal processing techniques, wavelet-based noise reduction has a major potential application to filter data, due to its excellent feature localization property. It reveals the information at a level of detail, which is not available with Fourier-based methods [4]. The so-called wavelet shrinkage employs nonlinear soft thresholding functions in the wavelet domain [6]. Also, the fast discrete wavelet signal transform (DWT) algorithms can fulfil time-critical needs. However, the selection of the appropriate wavelet and number of resolution levels, threshold function, etc. is still a challenging task. The point cloud may contain several objects, artifacts with different surface complexity thus the different parts need the appropriate shrinkage method. This paper introduces the fuzzy information measure for supporting the adaptive shrinkage procedure selection. The here described method conserns outlier removal and feature extraction. The complexity of an object is determined by the number of edges. In order to conserve precisely the shape that may contain the information of interest with simultaneously filtering out the noise, the objects with high complexity are processed with higher levels of resolution and appropriate threshold function. While the other parts can be smoothed. Furthermore, processing all of the points can be time-consuming, thus the fuzzy performance measure proposed for selecting the minimum amount of data that is necessary for obtaining the desired result. This serves also a basis for efficient data compression. The applicability and efficiency of this method have been validated via simulation results.

2 2.1

Point Cloud Processing with the Combination of Fuzzy Information Measure and Wavelets Principles of Discrete Wavelet Shrinkage

The point cloud contains points that are returned from the terrain objects, including ground, buildings, bridges, vehicles, trees, and other non-ground features. For many applications it is important to detect, separated, or removed the artifacts in order to extract the required information [17]. Many mathematical tools have been derived for other applications need the precise reconstruction of the artifacts or involve searching for the minimum of a continuous function or

Point Cloud Processing with the Combination of Fuzzy Information Measure

457

a sizing optimization process [11]. Thus, the pre-processing is almost important step which includes operations such as remove of systematic errors, filtering, feature detection and extraction, etc. The utilization and theory of wavelets is well established, for e.g. see [4]. The discrete wavelet transform analyses the signal at different frequency scales with different resolutions by reducing the signal into approximate and detail information. For removing noise, wavelet shrinkage employs nonlinear soft thresholding functions in the wavelet domain. The popularity of this nonparametric method is due to the excellent localization and feature extracting behavior. However several threshold estimators exist, it is still a challenging task to select the appropriate shrinkage method that fits to the type of signal and contaminating noise and further, is robust against impulse type noises [6]. The other major issue in noise reduction is minimizing the effects of extreme values or elements that deviate from the observation pattern (outliers) (see, for e.g. [2]). The first step of wavelet shrinkage is the decomposition of signal into the wavelet (detail and approximate) coefficients, as described in [9]. The basic idea behind wavelet shrinkage is setting the value of these coefficients with small magnitude to zero (hard thresholding), or setting their value determined by a lambda (λ) threshold level [7]. After, the reconstruction is carried out by performing the inverse discrete wavelet transform (IDWT). Generally, the shrinkage methods construct nonlinear threshold functions that relies on some statistical considerations. For instance, the smoothness-adaptive method (SureShrink) [7] is proposed to threshold each dyadic resolution level using the principle of Steins Unbiased Estimate of Risk [12], while the universal bound thresholding rule provides results with low computational complexity. The rule of the latter is defined, as follows [7], √ (1) ν = σM AD 2 log sj median(ω )

where σM AD = 0.6745 j denotes the absolute median deviation. The specific choice of the wavelet function, decomposition level, and thresholding rule allows to construct many different shrinkage procedures. 2.2

The Concept of Fuzzy Information Measure

The key issue of image processing, image understanding, data storage, information re-trieval, etc. is what carries the information in images or scenes. However, a wide range of possible concepts replies to this question, containing statistical approaches. Another, more sophisticated approach may also consider the elimination of the additive noises. This raises another question: What is noise? The main problem of answering this questions that noise is an ill-defined category, because noise is usually dependent on the situation and on the objective of the processing [15]. What is characteristic or useful information in one application can be noise in another one. Based on these consideration, the amount of information in an image or point cloud strongly related to the number and complexity of the objects in it [15]. The most characteristics of the objects are their boundaries, i.e. the edges that can be extracted. Since the boundary edges

458

A. Dineva et al.

of the objects contain the primary information about the object shape, these serve as base for classification, object recognition, etc. According to the concept of F uzzy Inf ormation M easure revealed in [14,15], also the point clouds content information can be represented by the characteristic features, like boundary edges. Thus, the amount of information in the 3D map is proportional to the number of characteristic boundary lines. 2.3

Fuzzy Information Measure for Feature Extraction

There are also some studies proposing edge detecting procedures for 3D point clouds [3]. Generally, the precise construction of a 3D edge map of the environment from remotely sensed data can be limited. The method proposed further extends our previously approach [5]. It has been shown that applying robust fitting on the wavelet coefficients can efficiently denoise the data. However, objects with high complexity need shrinking with higher resolution levels and performing the fitting with higher order polynomials for preserving the details. While, the other regions that contain unimportant or sparse information can be smoothed. For separating complex objects the number and distance of the most characteristic edges are counted that carries the information. For this reason, it is not necessary to construct the full edge map, just defining the areas where the density of these edges are high. For properly setting the parameters of the wavelet resolution and smoothing the fuzzy information measure, the approach supports this clustering task. As has been described, for e.g. in [3], the local neighborhood of the point is used to test whether a point lies on a potential edge or not. If there is a depth discontinuity in the lidar scan at the point is at a maximum, then that point is part of a contour edge. According to this concept, the fuzzy-edgeness is determined by the depth discontinuity. The discontinuity is determined in the wavelet domain. The μ(r) = 1 fuzzy membership value belongs to a region that carries the highest information. The fuzzy information measure of each region is determined as follows, 1 (2) μ(ri ) = (nrmax −nri ) 1+e in which nrmax denotes the number of edges in the rmax region that contains the maximum number of edges, while nri is the number of edges in the ith region. After, the areas with a high number of edges are at first decomposed with symlet wavelets using six levels. This step serves for the separation of the noise and the parts of the signal that carries the important information. After this, the robust fitting is performed on the approximate coefficients. Then, the iterative reweighting of the atoms is performed based on their residuals and using a trisquare function in a short running window, according to [3]. At last, the inverse discrete wavelet transform (IDWT) is performed. The same procedure holds for the regions with lower measure, but with applying daubechies wavelets with two levels of decomposition and for the fitting a bisquare function within a long sliding window.

Point Cloud Processing with the Combination of Fuzzy Information Measure

2.4

459

Fuzzy Performance Measure

In engineering practice, the compression ratio measures the relative decrease in complexity of data in a raw form comparing it to the complexity of the compressed form after processing. However, several measures have been applied [13]. In the presented approach the robust fitting performed on the wavelet coefficients can be time consuming if we calculate the weights and new values

Fig. 1. The raw test data (available at [1]) with additive white Gaussian noise (AWGN).

Fig. 2. Result of the denoising using wavelet shrinkage and fuzzy information measure. The raw data is available at [1].

460

A. Dineva et al.

for all the points. Once a point selected for evaluation it is not necessary for calculating again the weights in its vicinity. For selecting the points for the procedure the evaluated fuzzy information measure of the region can be used. In region ri the running window size determined by the dmax distance of the edges. In the regions mu(ri ) > 0.8 the 85% of the points are selected for evaluation, while within the regions 0.8 > mu(ri ) > 0.6 the 50% of the points are enough. These boundaries can be set according to experimental results.

3

Simulation Results

The proposed approach has been tested on a lidar data available at [1]. The data set contains 65536 points. In Fig. 1 the raw data can be seen that has been affected by noise. Figure 2 shows the result of the denoising with the combined approach. It can be seen, that the here described procedure ensures acceptable performance.

4

Conclusions

In this paper a novel concept has been proposed for denoising and feature extracting of remotely sensed point clouds. The here described method relies on the combination of fuzzy information measure and wavelets that allows precise feature extraction, denoising and additionally effective compression of the point cloud. The suitable setting of the applied wavelet and parameters have a great impact on the result and depends on the aim of preprocessing that usually is a challenge for the users. In order to address this problem the fuzzy information measure has been proposed that supports the adaptive setting of the parameters. However, the evaluation of all the points in a large data set is time consuming. For this reason, a fuzzy performance measure has been introduced, that can reduce the complexity of the procedure by reducing the number of points due to the separated regions fuzzy information measure values. Simulation results validate the efficiency and applicability of this method. Acknowledgements. This work has been sponsored by the Hungarian National Scientific Research Fund (OTKA 105846). This publication is also the partial result of the Research & Development Operational Programme for the project “Modernisation and Improvement of Technical Infrastructure for Research and Development of J. Selye University in the Fields of Nanotechnology and Intelligent Space”, ITMS 26210120042, co-funded by the European Regional Development Fund.

References 1. ISPRS test on extracting DEMs from point clouds: a comparison of existing automatic filters. http://www.itc.nl/isprswgiii-3/filtertest/ 2. Alvera-Azc´ arate, A.: Outlier detection in satellite data using spatial coherence. Remote Sens. Environ. 118, 84–91 (2012)

Point Cloud Processing with the Combination of Fuzzy Information Measure

461

3. Borges, P., Zlot, R., Bosse, M., Nuske, S., Tews, A.: Vision-based localiztaion using edge map extracted from 3D laser range data. In: IEEE International Conference on Robotics and Automation, pp. 4902–4909, May 3–8, Anchorage, Alaska, USA (2010) 4. Daubechies, I.: The wavelet transform, time frequency localizlocal and signal analysis. IEEE Trans. Inf. Theory 36, 961–1005 (1990) 5. Dineva, A., V´ arkonyi-K´ oczy, A.R., Tar, J.K.: Improved Denoising with Robust Fitting in the Wavelet Transform Domain. In: Technological Innovation for CloudBased Engineering System. IFIP Advances in Information and Communication Technology, vol. 450, pp. 179–187. Springer, Cham (2015) 6. Donoho, D.: Denoising by soft-thresholding. IEEE Trans. Inf. Theory 41(3), 613– 627 (1995) 7. Donoho, D., Johnstone, M.: Adapting to unknown smoothness via wavelet shrinkage. J. Am. Stat. Assoc. 90, 1200–1224 (1995) 8. Hamming, R.W.: Digital Filters, 3rd edn. Prentice-Hall, Englewood Cliffs (1989) 9. Mallat, S.: Multiresolution approximations and wavelet orthonormal bases of l2 . Trans. Am. Math. Soc. 315, 69–78 (1989) 10. Orfanidis, S.J.: Introduction to Signal Processing. Prentice-Hall Inc, Upper Saddle River (1995). ISBN 0-13-209172-0 11. Park, J.Y., Han, S.: Application of artificial bee colony algorithm to topology optimization for dynamic stiffness problems. Comput. Math. Appl. 66, 1879–1891 (2013) 12. Stein, C.M.: Estimation of the mean of a multivariate normal distribution. Ann. Stat. 9(6), 1135–1151 (1981) 13. Teolis, A.: Computational Signal Processing with Wavelets. Birkhauser, Boston (1998) 14. V´ arkonyi-K´ oczy, A.R., Hancsicska, S., Bukor, J.: Fuzzy information measure for improving HDR imaging. In: Shabbazova, S. (ed) In: Proceedings of the 4th World Conference on Soft Computing, WConSC 2014, pp. 491–497 (2014) 15. V´ arkonyi-K´ oczy, A.R., T´ oth, J.: A fuzzy information measure for image quality improvement. In: Nagatsu, M. (ed) In: Proceedings of the 14th International Conference on Global Research and Education in Intelligent Systems, Interacademia 2015, pp. 30–37 (2015) 16. Vaseghi, S.V.: Advanced Digital Signal Processing and Noise Reduction, 4th edn. Wiley, Hoboken (2008) 17. Vosselmann, G., Maas, H.G. (eds.): Airborne and Terrestrial Laser Scanning. Whittles Publishing, Dunbeath (2010)

Significance of Glottal Closure Instants Detection Algorithms in Vocal Emotion Conversion Susmitha Vekkot and Shikha Tripathi(&) Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Amrita University, Bengaluru, India {v_susmitha,t_shikha}@blr.amrita.edu

Abstract. The objective of this work is to explore the significance of efficient glottal activity detection for inter-emotion conversion. Performance of popular glottal epoch detection algorithms like Dynamic Projected Phase-Slope Algorithm (DYPSA), Speech Event Detection using Residual Excitation And a Mean-based Signal (SEDREAMS) and Zero Frequency Filtering (ZFF) are compared in the context of vocal emotion conversion. Existing conversion approaches deal with synthesis/conversion from neutral to different emotions. In this work, we have demonstrated the efficacy of determining the conversion parameters based on statistical values derived from multiple emotions and using them for inter-emotion conversion in Indian context. Pitch modification is effected by using transformation scales derived from both male and female speakers in IIT Kharagpur-Simulated Emotion Speech Corpus. Three archetypal emotions viz. anger, fear and happiness were generated using pitch and amplitude modification algorithm. Analysis of statistical parameters for pitch after conversion revealed that anger gives good subjective and objective similarity while characteristics of fear and happiness are most challenging to synthesise. Also, use of male voice for synthesis gave better intelligibility. Glottal activity detection by ZFF gave results with least error for median pitch. The results from this study indicated that for emotions with overlapping characteristics like surprise and happiness, inter-emotion conversion can be a better choice than conversion from neutral. Keywords: Glottal Closure Instants Inter-emotion conversion




DYPSA




SEDREAMS




ZFF




1 Introduction Emotions form a vital part of human communication. An emotion can be defined as that state of mind which conveys the non-linguistic information from the speaker. Emotions can be expressed through multiple modalities like facial expressions, hand gestures or speech. Vocal emotion forms the non-linguistic part of expression and has a range of variations depending on speaker gender, state of mind, physiology and language. Due to this, the recognition and analysis of human emotions has always been a challenging task. Vocal emotions form part of the para-linguistic information conveyed © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_40

Significance of Glottal Closure Instants Detection Algorithms

463

in the form of spoken language. With the advent of robotic communication technology, human-machine interaction has been gaining immense momentum during the last decade. Effective portrayal of emotions by robots requires that the parameters characterising the synthesis of emotions are estimated, analysed and manipulated at segmental, supra-segmental and sub-segmental levels. In Fants source-filter model of speech production, the source signal produced at lungs and manipulated by glottal constrictions is linearly filtered through the vocal tract. The operation can be compared to convolution using a glottal source and filter. The vocal tract shapes the characteristics of the vibrations emitted from the source. The resulting sound is emitted to the surrounding air through lips [1]. Glottal Closure Instants (GCIs) are defined as instances of significant excitation of resonances of vocal tract filter which corresponds to high energy in the glottal signal during voiced segment of speech and the onset points of frication in case of unvoiced speech [2]. Accurate detection of GCIs can aid in applications such as prosody modification, speech coding/compression and speaker recognition [3]. The essence of speech production starts from glottal flow and hence a reliable estimate of parameters contributing to emotion impression in speech has to start from accurate determination of dynamics of glottal flow. Modulation characteristics of airflow can be studied by analysing and comparing the glottal waves across multiple emotions. Different algorithms have been proposed for determination of GCIs of voiced speech. These algorithms follow different approaches. Majority of them use Linear Prediction residual (LPR) for epoch extraction which can be used as the best alternative for glottal wave/its derivative [4, 5]. Fewer techniques are available for directly estimating glottal instants from speech. Researchers have been interested in developing accurate methods for glottal activity detection from the last few years. One of the important algorithms for GCI detection is done by Naylor in [6], popularly known as Dynamic Projected Phase-Slope Algorithm (DYPSA). A breakthrough algorithm for accurate detection of GCIs using Zero Frequency Filtered Signal (ZFF) was proposed in [8, 9]. One of the prominent works on glottal activity detection was carried out by Sturme by using multiscale analysis through wavelet transforms [10]. The authors computed absolute maxima along various scales and generated a tree representation. The GCIs were located at the top of the tree branches. A critical evaluation of the various algorithms for detecting GA regions was done in [11]. Important algorithms like DYPSA [6], SEDREAMS [7], ZFF [8] and wavelet based Lines Of Maximum Amplitude (LOMA) [10] were compared using data from speakers of CMU-ARCTIC database. SEDREAMS was found to have better accuracy in terms of identification rates. Another useful technique for detecting GOIs is by estimating the Hilbert transform of the LP residual of speech [12]. Here, GOIs are estimated from difference EGG by picking secondary peaks. Performance evaluation showed an identification rate of 99.91%. In this paper we focus on the significance of glottal activity detection in conversion across multiple emotions. The rest of the paper is organised as follows: Sect. 2 describes the algorithms used to obtain the GCIs in the paper, Sect. 3 describes the implementation of inter-emotion conversion using the various epoch detection algorithms while Sect. 4 describes the results and objective evaluation using statistical values and comparison of error percentage. Finally, the paper is concluded in Sect. 5 with insights into the challenges involved and future scope of work.

464

S. Vekkot and S. Tripathi

2 Algorithms for Glottal Epoch Detection This section throws light on the algorithms used for epoch detection from speech signals. 2.1

Dynamic Projected Phase-Slope Algorithm (DYPSA)

The algorithm described in [6] uses a modification of [13] in which Group Delay function is used to determine instants of significant excitation of glottis. As the name implies, DYPSA algorithm works on the slope of unwrapped phase of Short-Time Fourier transformed Linear Prediction (LP) residual, which is known as phase slope in Group Delay Algorithm [6]. The peak GCI candidates are identified from the positive going and projected zero-crossings of the above function. Finally, a dynamic programming approach is used to select the optimum candidates for glottal closure by minimising an objective function. 2.2

Speech Event Detection Using Residual Excitation and a Mean-Based Signal (SEDREAMS)

Speech Event Detection using Residual Excitation And a Mean-based Signal (SEDREAMS) uses a mean-based signal for epoch detection [7]. This signal is obtained as given in Eq. 1: N X 1 ym ¼ wðmÞ x ðn þ mÞ ð1Þ 2N þ 1 m¼N where x(n) is speech signal and w(m) is the window used. The GCI intervals extracted from the above signal may not give accurate detection. Hence mean based minima are found out and intervals between 2 successive minima (starting from mean based minima till 0.35 times local pitch period) are taken to represent the relative positions of GCIs [7]. For an accurate estimation, the local maxima within the above interval is found out which refines the actual GCI positions. 2.3

Zero Frequency Filtering

ZFF algorithm locates epochs in speech by computing the zero frequency filtered signal [8]. Initially, the speech signal is pre-processed to remove any irregularities in recording: dðnÞ ¼ sðnÞ  sðn  1Þ ð2Þ d(n) is known as difference speech. This is filtered through a cascade of two ideal zero-frequency resonators:

Significance of Glottal Closure Instants Detection Algorithms

yðnÞ ¼

4 X

ak yðn  kÞ þ dðnÞ

k¼1

465

ð3Þ

fa1 ; a2 ; a3 ; a4 g ¼ f4; 6; 4; 1g For removing the trend in y(n), the local mean subtraction is done at each sample to get the zero frequency filtered signal, yZFF(n). yZFF ðnÞ ¼ yðnÞ 

N X 1 yðn þ mÞ 2N þ 1 m¼N

ð4Þ

where 2N + 1 is the window length used. Epoch locations can be estimated by locating the positive zero-crossings of yZFF(n).

3 Implementation 3.1

Database Used

The database used for experimentation is IITKGP-SESC (Indian Institute of Technology Kharagpur Simulated Emotion Speech Corpus) which is recorded using 10 (5 male and 5 female) experienced professional artists from All India Radio (AIR) Vijayawada, India. For analysing the emotions 15 emotionally neutral Telugu sentences are considered in which each artist speaks with 8 emotions viz. neutral, anger, happy, fear, compassion, sarcasm, surprise and disgust. Each emotion has 1500 utterances, thus making the total utterances to 12000, lasting for around 7 h [14]. The speech signal was sampled at 16 kHz. For the current study, we have considered 4 basic emotions in IIT-KGP database viz. neutral, anger, happy and fear from male and female artists. For each of the emotions, statistical values i.e. median, mean and standard deviation of pitch are calculated from 10 different utterances. For each speech sample, GCIs are located using each of the above discussed algorithms and used for inter-emotion conversion using the statistical values computed. Inter-emotion conversion among the above 4 emotions is carried out and results are tabulated. The statistical values for pitch of the emotions considered in this paper for male and female speakers are given in Table 1. 3.2

Pitch Transformation Scales

Using the statistical values obtained above, the transformation scales for conversion across different emotions have been obtained using Eq. 5: Pitchfactorð%Þ ¼

T arg etpitch  Originalpitch  100 Originalpitch

ð5Þ

For analysis, the percentage change is chosen by taking the median value of each of the pitch parameters in the database. This is done to overcome the changes in the mean

466

S. Vekkot and S. Tripathi Table 1. Gender-wise statistical pitch values obtained for IIT-KGP SESC database Gender Emotion Median Male Neutral 187.38 Anger 272.72 Fear 249.76 Happy 219.56 Surprise 252.23 Female Neutral 303.69 Anger 370.50 Fear 309.62 Happy 316.5 Surprise 355.99

Mean 186.83 264.64 248.67 216.87 251.36 310.19 381.52 310.6 338.07 361.34

Standard deviation 40.09 50.77 28.27 42.83 56.85 80.4 95.96 52.72 94.8 85.85

value due to few irregularities in parameter values. Median of a data distribution gives a more accurate estimation as it does not take into account the irregularities. Also, the median and mean differ by less than 5% [15]. The percentage change to be incorporated in original emotions for conversion in each case is given in Table 2. Applying the pitch transformation scales below, the variations in pitch has been effected corresponding to each emotion. After pitch scaling, the resulting speech sample is subjected to amplitude scaling by multiplying the intermediate signal with the amplitude tier of the target emotion. Thus pitch and amplitude scaling for each emotion is performed. The statistical parameters are recalculated after conversion. The results are compared with the median values for target emotion as we have derived the transform scales based on median value. 3.3

Inter-emotion Conversion

Implementation of emotional conversion is carried out in this paper as a two-stage process. The first stage involves incorporating the transformation scales obtained from Table 2 in pitch modification algorithm. The first stage of the conversion process makes use of dynamic pitch modification algorithm depicted by Govind et al. [16]. This method has been chosen based on the efficiency in preserving the prosodic variations for high arousal emotions. Stage I 1. Estimate the instants of significant excitation/glottal closure using the algorithm chosen for GCI detection. 2. Get the epoch intervals as difference between successive GCIs. 3. Modification factor, mq for every qth interval is expressed as an integer ratio Aq/Bq. The initial value for modification factor is the transformation scale value derived from Table 2. 4. mq corresponding to each emotion is derived based on transform scale Eq. 5. 5. New epoch instants are obtained by equally dividing the qth and q + Ath location into Bq intervals.

Significance of Glottal Closure Instants Detection Algorithms Table 2. Transformation scales obtained for inter-emotion conversion Gender Original emotion Target emotion Male Neutral Anger Fear Happy Surprise Anger Neutral Fear Happy Surprise Fear Neutral Anger Happy Surprise Happy Neutral Anger Fear Surprise Surprise Neutral Anger Fear Happy Female Neutral Anger Fear Happy Surprise Anger Neutral Fear Happy Surprise Fear Neutral Anger Happy Surprise Happy Neutral Anger Fear Surprise Surprise Neutral Anger Fear Happy

Change (%) 45.54 33.28 17.17 34.60 −31.29 −8.43 −19.49 −7.52 −24.97 9.2 −12.1 0.99 −14.65 24.22 13.76 14.88 −25.71 8.13 −0.99 −12.95 23.00 1.95 4.20 17.22 −18.03 −16.40 −14.57 −3.92 1.91 19.67 2.22 14.98 −4.05 17.06 −2.18 12.47 −14.69 4.08 −13.03 −11.09

467

468

S. Vekkot and S. Tripathi

6. The modified epoch locations are derived from modified instants computed in step 5 of the algorithm. 7. Similarly, the modified instants for each successive interval are calculated by incrementing qth location value by 1. 8. Repeat steps 3–7 until end location to generate new instants sequence. 9. Modified instants are concatenated and new pitch anchor points generated. 10. Speech waveform is reconstructed by copying the speech samples existing in each epoch interval to the adjacent modified epoch interval in the modified instants sequence obtained [17]. Figure 1 plots the target and synthesized happy speech with their pitch contours and spectrograms after Stage I (pitch scaling using transformation scale obtained in each case) of the algorithm.

Fig. 1. Results obtained after pitch modification for neutral to happy conversion with GCI estimation using different algorithms. (a) Original happy speech segment, pitch contour and spectrogram. (b) Converted happy, pitch contour and spectrogram with GCI estimation using DYPSA. (c) Converted happy speech, pitch contour and spectrogram with GCI estimation using SEDREAMS. (d) Converted happy speech, pitch contour and spectrogram with GCI estimation using ZFF

Significance of Glottal Closure Instants Detection Algorithms

469

The second stage involves intensity modification by imposition of amplitude changes of target emotion into the pitch modified signal from the first stage. The algorithmic steps involved in conversion are listed below: Stage II 1. The pitch modified speech signal is saved as.wav file and extracted for further processing. 2. The target emotion speech file is processed to separate the intensity tier. 3. The amplitude variations in target emotion are imposed on pitch modified speech by multiplying the extracted amplitude tier with the pitch modified speech signal.

4 Results and Objective Evaluation The statistical parameters are again calculated after conversion. The results are compared with the median values for original emotion as we have derived the transform scales based on median value. Table 3. Comparison between statistical values for synthesised emotion and target emotion based on GCI algorithms GCI algorithm

Gender Statistical value

Source emotion Neutral Fear Happy Surprise Anger

Target values

Target emotion

DYPSA

Male

282.77 293.44 58.83 289.54 294.3 83.50 212.87 200.46 56.78 236.15 234.68 78.25 222.08 209.4 64.26 298.27 313.74 79.64 281.17 265.24 47.96 331.86 326.53 62.80

– – – – – – 193.26 197.26 45.8 245.20 294.05 115.4 213.55 208.44 22.18 300.00 295.00 92.48 – – – – – –

272.7 264.64 50.77 357.61 351.68 75.84 249.76 248.67 28.27 290.15 284.79 51.05 219.56 216.86 42.83 305.33 327.57 72.3 272.70 264.64 50.77 357.61 351.68 75.84

Anger

Female

Male

Female

Male

Female

SEDREAMS Male

Female

Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev

278.6 272.21 37.87 371.80 367.37 48.78 – – – – – – 227.36 213.87 38.97 233.68 234.94 71.75 312.85 305.00 24.50 398.38 392.21 22.20

276.17 252.75 62.63 343.67 359.42 58.94 215.59 207.83 30.45 244.72 239.18 32.00 – – – – – – 291.37 286.82 31.46 332.46 334.19 86.32

299.42 265.53 78.92 398.95 363.23 84.59 214.88 188.13 49.09 297.73 284.2 60.04 255.70 254.17 38.46 310.00 304.40 38.56 299.40 298.33 51.86 467.2 467.56 10.75

Fear

Happy

Anger

(continued)

470

S. Vekkot and S. Tripathi Table 3. (continued)

GCI algorithm

Gender Statistical value Male

Female

Male

Female

ZFF

Male

Female

Male

Female

Male

Female

Source emotion Neutral Fear Happy Surprise Anger

Target values

Target emotion

Median

216.78 –

249.76

Fear

Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev Median Mean Std. Dev

227.35 36.52 309.82 296.54 62.03 228.13 237.36 39.00 309.86 330.85 89.00 263.86 268.33 47.87 372.15 361.91 64.27 223.62 232.98 37.48 271.68 265.67 37.12 224.90 233.23 39.86 301.60 293.08 60.50

– – – – – 219.02 210.34 29.30 338.50 328.57 74.76 272.87 269.92 33.97 373.88 377.50 55.70 – – – – – – 221.30 219.74 18.81 281.10 297.46 42.12

235.42 240.48 38.90 252.65 250.22 11.28 – – – – – – 270.59 261.30 36.36 351.78 371.05 47.34 230.60 224.68 29.59 320.90 309.97 94.00 – – – – – –

233.54 43.00 303.47 291.24 46.48 238.61 235.36 37.73 310.24 306.89 20.77 236.24 246.00 52.77 279.82 323.48 89.97 234.74 235.80 43.60 298.12 301.77 78.35 242.76 225.99 45.20 302.49 316.09 67.77

218.16 31.60 203.58 238.99 92.53 204.4 195.27 40.75 193.00 216.3 70.04 – – – – – – 220.44 215.23 32.98 248.00 245.82 60.62 202.97 202.27 19.62 260.70 278.07 44.96

248.67 28.27 290.15 284.79 51.05 219.56 216.86 42.83 305.33 327.57 72.3 272.70 264.64 50.77 357.61 351.68 75.84 249.76 248.67 28.27 290.15 284.79 51.05 219.56 216.86 42.83 305.33 327.57 72.3

Happy

Anger

Fear

Happy

The results for converted (synthesized) emotion are tabulated in Table 3 and are compared with target emotion in each case. The percentage error rates have been calculated for each GCI algorithm based inter-emotion conversion and the results have been listed in Table 4. A graphical comparison between the algorithms has been drawn in Fig. 2. As we have used the scaling factor taking into account the median values, a comparison has been drawn based on median values.

Significance of Glottal Closure Instants Detection Algorithms

471

Table 4. Comparison of error (%) for inter-emotion conversion based on GCI algorithms Gender GCI algorithm Error (%) when target emotion is: Anger Fear Happy Male DYPSA 4.23 16.25 4.60 SEDREAMS 8.60 6.85 1.36 ZFF 4.33 8.97 1.56 Female DYPSA 1.85 11.78 6.49 SEDREAMS 6.95 7.85 5.70 ZFF 3.69 1.89 6.18

Fig. 2. Comparison of various GCI algorithms based on percentage error. The values are calculated based on median of anger, fear and happy emotions for both male and female speakers

From the experiments, some observations have been made: • Median value forms a better estimate for probabilistic study of various emotions as it is free from irregularities. • Conversion efficiency is higher for anger target in both neutral-emotion and inter-emotion cases. This is because anger is mainly categorized by higher pitch and intensity. The variations done as part of this work also focussed on pitch and intensity scaling. Anger gave best results for perception also. • Most deviation in results was obtained for conversion to/from fear emotion. Fear emotion is characterised with large variance and lower energy with the addition of a panic or anxiety component, thus giving rise to a breathy quality to voice. Mere pitch, duration and amplitude variations are unable to incorporate the harmonics and voice quality of fear.

472

S. Vekkot and S. Tripathi

• Conversion to/from happy also yielded average perception results. But the pitch contour of the synthesised emotion was found to be matching with original happy to some extent as shown by Fig. 1. Though the algorithm succeeded in bringing the pitch and intensity to the required level, it still failed to bring in the smooth rippling kind of inflexions normally found in happy speech. • From the error comparison Table 4, the minimum percentage error is obtained for emotion conversion in case of male speaker. As far as inter-emotion conversion is concerned, though the trends are fluctuating for the various algorithms we can see that an optimal choice of ZFF for epoch estimation gives acceptable error values. Also, the conversion error is maximum for fear case in both male and female speakers which is rightly correlated by experimental findings and perception. • The average computation time required for execution of conversion algorithm from neutral to emotion and between emotions were computed for each GCI algorithm and it was found that they are closely similar (around 220–260 µs). This shows that no restriction exists while choosing the mode of conversion (neutral emotion/emotion - emotion) and either can be chosen depending on interrelationship between source and target. For instance, conversion from surprise to happy yielded good objective and perceptive results in all cases (Table 3). This is due to the fact that surprise is depicted in the database as “pleasant”. • We found that the conversion efficiency depends on expressivity of database to a large extent. Emotions like happy and fear can be better distinguished from angry and sadness if the speaker displays the emotion with characteristic precision. In some cases, the enacted database fails to bring in the complex inundations and voice modulations of emotions while spontaneous recording can bring in a much perceivable quality to the target.

5 Conclusions The experiments on inter-emotion conversion gave acceptable results for pitch variations. Conversion to anger yielded best objective and perceptive results while fear and happiness conversions failed to bring in the dynamics of emotion involved. For emotions like happiness and surprise, the synthesised emotion was better perceived when converted from each other than from neutral emotion. However, though the computation time required for inter-emotion conversion using various GCI detection algorithms were similar, ZFF gave best results in terms of conversion error rate (%). Studies proved that in addition to incorporating variations in prosody, determining the voice quality parameters in emotion conversion and embedding these at proper instances are instrumental in making the target more expressive and intelligible. The challenge lies in suitably inserting the parameters thus determined without distorting the spectral characteristics of normal speech and further, making a generalized framework for emotion conversion which is database, speaker and language independent.

Significance of Glottal Closure Instants Detection Algorithms

473

Acknowledgements. We would like to thank Dr. Govind. D, Amrita Vishwa Vidyapeetham, Coimbatore for providing us with IIT-KGP SESC database used in this research.

References 1. Fant, G.: Acoustic Theory of Speech Production. Mouton, The Hague (1960) 2. Murthy, K.S.R., Yegnanarayana, B.: Epoch extraction from speech signals. IEEE Trans. Audio Speech Lang. Proc. 16(8), 1602–1614 (2008) 3. Rao, K.S., Yegnanarayana, B.: Prosody modification using instants of significant excitation. IEEE Trans. Audio Speech Lang. Proc. 14, 972–980 (2006) 4. Ananthapadmanabha, T.V., Yegnanarayana, B.: Epoch extraction from linear prediction residual for identification of closed glottis interval. IEEE Trans. Acoust. Speech Signal Proc. ASSP-27, 309–319 (1979) 5. Prasanna, S.R.M., Govind, D.: Analysis of excitation source information in emotional speech. In: Proceedings of the Interspeech, Makuhari, Chiba, pp. 781–784 (2010) 6. Naylor, P., Kounoudes, A., Gudnason, J., Brookes, M.: Estimation of glottal closure instants in voiced speech using the dypsa algorithm. IEEE Trans. Audio Speech Lang. Proc. 15(1), 34–43 (2007) 7. Drugman, T., Dutoit, T.: Glottal closure and opening instant detection from speech signals. In: Proceedings of Interspeech, 6–10 September 2009, pp. 2891–2894 (2009) 8. Yegnanarayana, B., Murthy, K.S.R.: Event-based instantaneous fundamental frequency estimation from speech signals. IEEE Trans. Audio Speech Lang. Proc. 17(4), 614–624 (2009) 9. Murthy, K.S.R., Yegnanarayana, B., Joseph, A.: Characterisation of glottal activity from speech signals. IEEE Signal Process. Lett. 16(6), 469–472 (2009) 10. Sturmel. N., d’Alessandro, C., Rigaud, F.: Glottal closure instant detection using lines of maximum amplitudes (LOMA) of the wavelet transform. In: Proceedings of ICASSP, pp. 4517–4520 (2009) 11. Cabral, J.P., Kane, J., Gobl, C., Carson-Berndsen, J.: Evaluation of glottal epoch detection algorithms on different voice types. In: Proceedings of the Interspeech, Florence, Italy, pp. 1989–1992 (2011) 12. Ramesh, K., Prasanna, S.R.M., Govind, D.: Detection of glottal opening instants using hilbert envelope. In: Proceedings of the Interspeech, Lyon, France, pp. 44–48 (2013) 13. Smits, R., Yegnanarayana, B.: Determination of instants of significant excitation in speech using group delay function. IEEE Trans. Audio Speech Lang. Proc. 3, 325–333 (1995) 14. Koolagudi, S.G., Maity, S., Vuppala, A.K., Chakrabarti, S., Rao, K.S.: IITKGP-SESC: speech database for emotion analysis. In: Proceedings of IC3, Noida, pp. 485–492 (2009) 15. Akanksh, B., Vekkot, S., Tripathi, S.: Inter-conversion of emotions in speech using TDPSOLA. In: Proceedings of the Advances in Signal Processing and Intelligent Recognition Systems, SIRS 2015, pp. 367–378. Springer, Trivandrum (2015) 16. Govind, D., Joy, T.T.: Improving the flexibility of dynamic prosody modification using instants of significant excitation. Circ. Syst. Sig. Process. 35(7), 2518–2543 (2016). doi:10. 1007/s00034-015-0159-5 17. Govind, D., Prasanna, S.R.M.: Dynamic prosody modification using zero frequency filtered signal. Int. J. Speech Technol. 16(1), 41–54 (2013). doi:10.1007/s10772-012-9155-3

Analysis of Image Compression Approaches Using Wavelet Transform and Kohonen’s Network Mourad Rahali(&), Habiba Loukil, and Mohamed Salim Bouhlel Sciences and Technologies of Image and Telecommunications, High Institute of Biotechnology, University of Sfax, Sfax, Tunisia [email protected], [email protected], [email protected]

Abstract. Since digital images require a large space on the storage devices and the network bandwidth, many compression methods have been used to solve this problem. Actually, these methods have, more or less, good results in terms of compression ratio and the quality of the reconstructed images. There are two main types of compression: the lossless compression which is based on the scalar quantization and the lossy compression which rests on the vector quantization. Among the vector quantization algorithms, we can cite the Kohonen’s network. To improve the compression result, we add a pre-processing phase. This phase is performed on the image before applying the Kohonen’s network of compression. Such a phase is the wavelet transform. Indeed, this paper is meant to study and model an approach to image compression by using the wavelet transform and Kohonen’s network. The compression settings for the approach to the model are based on the quality metrics rwPSNR and MSSIM. Keywords: Image compression Learning algorithm


Kohonen’s networks
Wavelet transform


1 Introduction Digital images pose a problem of transmission time and storage volume. The compression techniques can reduce the space needed to represent a certain amount of information. The compression techniques are divided into two main categories. First, the lossy compression in which some of the information in the original image is lost. Second, the lossless compression exploits the information redundancy in the image to reduce its size. The lossy compression methods [1] are more likely to achieve higher compression ratio than those obtained by the lossless methods [2, 5]. The methods of image compression by neural networks [3] yield acceptable results. Yet, these methods have a limit on the compression ratio and the reconstructed image quality. To improve the reconstructed image quality, we combine the discrete wavelet transform (DWT) [11] and the quantization by Kohonen’s networks [4]. Thereafter, we use the Huffman coding to encode the quantized values [6, 14]. In this paper, we are interested in the study of an approach to image compression through the use of the wavelet transform and Kohonen’s network. We will, in particular, detail the learning © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_41

Analysis of Image Compression Approaches Using Wavelet Transform

475

process of image compression and evaluate the compression result with a new quality metric rwPSNR “relative weighted PSNR”. To develop and improve the assessment, we will use another quality metric; namely, the MSSIM “Means Structural SIMilarity”. Next, we test the image compression by using the wavelet and Kohonen’s network together. Lastly, we make a comparison by using Kohonen’s network only without the wavelet transform and the second comparison between the proposed approach and various compression methods.

2 Proposed Approach 2.1

Image Compression

Image compression is carried out through the following steps: • Apply a wavelet transform [9, 22] to an original image depending on the decomposition level and the wavelet type. • Decompose the image into blocks according to a block size (for example 2  2, 4  4, 8  8 or 16  16). • Search the codebook for each block and the code word with a minimum distance from the block. The index of the selected word is added to the index vector that represents the compressed image. • Code the index vector by a Huffman coding [14]. • Save the index vectors coded for use during decompression. The Fig. 1 depicts the steps of compression.

Original Image

Wavelet transform

Kohonen

Huffman coding

Compressed Image

Fig. 1. Image compression steps

2.2

Learning Phase

In fact, learning [7, 8] is deemed to be a very important step to compress images by the neural network. The goal is to construct codebooks to be used during compression. The learning process is described in Fig. 2. The first step of learning is the wavelet transform of an original image to obtain four sub-images: an approximation image and three detail-images (Fig. 4) in different resolutions depending on the decomposition level and the wavelet choice. The second step is to decompose the four sub-images in blocks according to the block sizes (2  2, 4  4, 8  8 or 16  16). The blocks are arranged in linear vectors to be presented in the Self-Organizing Map (SOM) [4, 16] one after the other. The third step is to adjust the weight-coupling according to an index vector. The weights obtained at the end of learning represent the codebook which will be used for compression. The codebook obtained depends on several settings such as the choice of the learning image, the type

476

M. Rahali et al.

of the wavelet transform, the decomposition level, the block size and the size of the self-organizing map. Therefore, we should create several codebooks to improve the compression.

Original image

Wavelet Transform

Decompose image

Codebook SOM

Fig. 2. Learning phase

2.3

Image Decompression

Image decompression is realized throughout these steps: • Replace each code by the corresponding index to obtain the index vector. This is the decoding step. • Find the three detail-images and the approximation image by replacing each element of the index vector by the corresponding block in the codebook. In order to improve the reconstructed image quality, in our approach, we keep the approximation image un-indexed by the self organization map. • The inverse transform is applied to the sub-images obtained after de-quantization to display the reconstructed image. The Fig. 3 described the decompression steps.

Compressed Image

Decoding

Dequantization

Inverse transform

Reconstructed Image

Fig. 3. Image decompression steps

The same codebook is used during both compression and decompression. 2.4

Kohonen’s Network Algorithm

Kohonen’s network algorithm [4, 10] follows these steps: • Find the winning neuron of the competition

Analysis of Image Compression Approaches Using Wavelet Transform

dðX; wc Þ  dðX; wi Þ; 8i 6¼ c

477

ð1Þ

where, X is input vector, wc is weight vector of the winning neuron c and wi is weight vector of the neuron i. • Update weight wi wi ðt þ 1Þ ¼ wi ðtÞ þ hðc; i; tÞ  ½X  wi ðtÞ

ð2Þ

where, wi is the weight vector of the neuron i in instant t and h is a function defined by:
hðc; i; tÞ ¼

aðtÞ; i 2 Nðc; tÞ with aðtÞ 2 ½0; 1 0; else if

ð3Þ

The function h defined the extent of the correction to the winning neuron c and its neighborhood. In instant t, the neighbors of winning neuron c are determined by the function Nðc; tÞ. The final neighbors of a neuron consist of the neuron itself. The function hðc; i; tÞ assigns the same correction aðtÞ for all neurons belonging to the neighbors of the winning neuron at instant t [18]. 2.5

Image Pretreatment Using Wavelet Transform

The two-dimension-wavelet transform [9, 10] is adopted in our approach. Figure 4 shows the image division into sub-images for the case of seven-band decomposition. The sub-image LL2 represents the lowest frequency of the original image. As a matter of fact, the restoration of the wavelet coefficients in the sub-image LL2 will directly affect the image quality. The sub-images HH1, LH1, HL1, HH2, LH2, and HL2 contain detail information of the edge, the outline and the vein of the image at different decomposition layers. Concretely, the sub-image HL2 and the sub-image HL1 denote the image-coefficient at the vertical edge after the first and the second layers wavelet decomposition. The sub-image LH2 and the sub-image LH1 indicate the image coefficients at the horizontal edge. The sub-image HH1 and the sub-image HH2 signify the image coefficients on the cross edge. LL2

HL1 HL1

LH1

LH1

HH1

HH1

Fig. 4. Decomposition on the frequency by wavelet

478

M. Rahali et al.

3 Objective Assessments: The Quality Index To improve our method, we use tow quality metric: rwPSNR and MSSIM. 3.1

The Relative Weighted Peak Signal to Noise Ratio rwPSNR

The PSNR quantifies an intensity of the distortion. It does not adjust to the dynamic characteristics of the image. Indeed, the deterioration is more visible in less textured zones (weak variance) and is less visible in more textured zones (stronger variance) [20]. Accordingly, we take the variance of the picture into consideration. Hence, it increases when the variance is high and decreases in the opposite case. We will have a new definition of the MSE. Let X = {xij | I = 1, …, M; j = 1, …, N}and Y = {yij | I = 1, …, M; j = 1, …, N} be the original image and the test image, respectively. The wMSE is given as:   2 1 X N 1   X xm;n  ym;n  1 M wMSE ¼ MN m¼0 n¼0 1 þ Var ðM; N Þ

ð4Þ

Where Var(M, N) is the test image variance. The human eyes do not have an equal sensitivity across the different intensities. In fact, there is a threshold of sensitivity that must be exceeded before an increase of the intensity so that it can be detected. So, as a complement to the wPSNR, we introduce our rwPSNR [12] “relative weighted PSNR” which takes account of the relative difference of the image-gray levels because the noticeable difference of the two stimuli is roughly proportional to the intensity of the stimulus. Actually, an error between two pixels of two images can not translate the same error deviation between two pixels of two other images with the same intensity difference. Indeed, if the intensity difference (10) between the pixels is 10 and 20, it remains numerically the same as that between a pair of pixel values 110 and 120. However, the perception differs on the visual plan. In the first case, the error is quantified at 100% (20 to 10). But, in the second case, the error is quantifiable at 10% (120-110). Therefore, one has to think about the necessity of introducing the relative difference notion in the calculation of the wPSNR from which rwPSNR is derived. So, we have a new definition of the MSE noted as rwMSE “relative weighted Mean Square Error” which takes account of the variance and the image intensity. Our rwMSE is defined as follows: Let X = {x | i = 1, …, M; j = 1, …, N} and Y = {y | i = 1, …, M; j = 1, …, N} respectively be the original image and the test image. The rwMSE is given as: 2 1 X N 1  X 1 M jðx  yÞ=ðx þ yÞj rwMSE ¼ 2 MN m¼0 n¼0 1 þ Var ðM; N Þ The expression of our relative weighted peak signal to noise ratio is given by:  2  xmax rwPSNR ¼ 10  log10 rwMSE

ð5Þ

ð6Þ

Analysis of Image Compression Approaches Using Wavelet Transform

3.2

479

The Structural Similarity Means MSSIM

The new image quality measurement design is based on the assumption that the human visual system is highly adapted to extract the structural information of the visual field. The measurement of the structural information change can provide a good approximation of the distortion of the perceived image. The error sensitivity approach estimates the perceived errors to quantify the image degradation [21]; whereas, the new philosophy considers the degradation of the image as the perceived changes in the structural information. The luminosity of the surface of the observed object is the product of illumination and reflection. But the structures of the objects in the scene are independent of illumination [23]. Therefore, to explore the structural information in an image, we have to eliminate the influence of illumination. Therefore, the structural information in an image is defined as the attributes that represent the structures of the objects. Since luminosity and contrast may vary across the scene, we use luminosity and the local contrasts in our definition. The system breaks the task of similarity measurement into three comparisons: Luminosity Lðx; yÞ, Contrast Cðx; yÞ and Structure Sðx; yÞ. The combination of the three comparisons determines the structural similarity index (SSIM) [13]. SSIMðx; yÞ ¼

ð2lx ly þ C1 Þð2rxy þ C2 Þ þ l2y þ C1Þðr2x þ r2y þ C2 Þ

ðl2x

ð7Þ

When applying, only one total quality measurement of the whole image is required; whence, a means SSIM index (MSSIM) to assess the overall quality of the image is determined. MSSIMðX; YÞ ¼

M 1X SSIMðxi ; yi Þ M i¼1

ð8Þ

4 Subjective Assessments The objective assessment is insufficient to assess the visual quality of the compression methods. The subjective assessment analyzes the degradation of visual quality of images. Also, the human eye can judge the compression quality to compare the result between compression methods [19]. In our work, we compare tow compression methods with and without wavelet transform. 4.1

Quantitative Assessments of Results

In our work, we change the compression parameters: the decomposition level (j) of the wavelet, the input block size (BS), the wavelet type (haar, Coiflets, Daubechies, Symlets, …) and the size of the self organizing map (SOM). To evaluate the performance of our approach in image compression, we use the following measures: bits per pixel (Nbpp), the means square error (MSE), the relative weighted peak signal to the noise ratio (rwPSNR) and the structural similarity means (MSSIM) and we compare our approach to image compression without using the wavelet transform (Tables 1 and 2).

480

M. Rahali et al. Table 1. With wavelet transform Image Lena

Nbpp rwPSNR MSE

4.60 4.17 3.02 1.94 0.59 0.34 Cameraman 4.71 3.90 2.79 2.32 1.82 1.29 0.57 Barbara 4.14 2.96 2.42 1.83 1.36 0.59 0.21

66.98 66.87 62.94 61.06 60.76 58.06 61.75 61.47 60.04 59.82 59.20 58.08 57.72 62.82 61.18 60.49 60.13 57.60 57.68 56.42

41.60 43.43 66.88 175.21 208.31 502.25 90.08 93.58 114.65 120.43 135.9 385.75 403.24 155.29 204.50 246.09 258.69 394.86 469.14 624.43

MSSIM Parameters 0.954 0.946 0.933 0.785 0.732 0.526 0.932 0.928 0.899 0.892 0.859 0.724 0.664 0.899 0.857 0.813 0.770 0.650 0.526 0.369

J J J J J J J J J J J J J J J J J J J J

= = = = = = = = = = = = = = = = = = = =

1; 1; 1; 2; 2; 3; 1; 1, 1, 1, 1, 2, 2, 1, 1, 1, 1, 2, 2, 3,

BS BS BS BS BS BS BS BS BS BS BS BS BS BS BS BS BS BS BS BS

= = = = = = = = = = = = = = = = = = = =

4; SOM = 256 4; SOM = 64 16; SOM = 256 16; SOM = 256 256; SOM = 16 64; SOM = 16 4; SOM = 256 4, SOM = 64 16, SOM = 256 16, SOM = 64 16, SOM = 4 16, SOM = 64 256, SOM = 64 4, SOM = 64 16, SOM = 256 4, SOM = 4 256, SOM = 4 16, SOM = 64 64, SOM = 4 256, SOM = 4

Table 2. Without wavelet transform Image Lena

Nbpp rwPSNR MSE

3.92 2.54 1.91 1.66 0.82 0.31 Cameraman 3.72 2.23 2.01 1.39 0.75 0.42 0.39 Barbara 3.23 2.99 2.36 1.56 0.75 0.47 0.33

63.45 61.14 60.70 59.73 61.03 57.47 60.43 59.86 56.04 57.21 57.65 55.79 55.65 61.53 60.09 57.56 56.61 58.62 55.61 54.83

92.43 133.26 151.82 175.75 248.99 566.33 120.62 177.74 253.46 211.84 415.16 454.63 530.23 156.83 181.62 265.42 316.59 366.85 507.30 563.21

MSSIM Parameters 0.938 0.869 0.845 0.753 0.738 0.486 0.898 0.863 0.833 0.739 0.703 0.661 0.502 0.865 0.837 0.792 0.687 0.643 0.541 0.364

BS BS BS BS BS BS BS BS BS BS BS BS BS BS BS BS BS BS BS BS

= = = = = = = = = = = = = = = = = = = =

4; SOM = 256 4; SOM = 64 16; SOM = 64 16; SOM = 16 64; SOM = 16 256; SOM = 16 4; SOM = 256 4; SOM = 64 16; SOM = 64 16; SOM = 16 64; SOM = 16 256; SOM = 64 256; SOM = 16 4; SOM = 256 4; SOM = 64 16; SOM = 64 16; SOM = 16 64; SOM = 16 256; SOM = 64 256; SOM = 16

Analysis of Image Compression Approaches Using Wavelet Transform

481

Table 3 shows an objective assessment of the relative weighted peak signal to the noise ratio (rwPSNR) depending on the number of bits per pixel (Nbpp). The blue curve represents the rwPSNR of the compressed images using the discrete wavelet transform and Kohonen’s network. The green curve represents the rwPSNR of the compressed images using only Kohonen’s network without the wavelet transform. Thus, we see very well that the image quality calculated by our metric (rwPSNR) has been improved by the compression approach using the wavelet transform compared to the other approach without wavelet transform. Table 4 shows an objective assessment of the structural similarity means (MSSIM) depending on the number of bits per pixel (Nbpp). The blue curve represents the MSSIM of the compressed images using the discrete wavelet transform and Kohonen’s network. The green curve represents the MSSIM of the compressed images using only Kohonen’s network without the wavelet transform.

Table 3. Curves of rwPSNR = f(Nbpp)

Table 4. Curves of MSSIM = f(Nbpp)

482

4.2

M. Rahali et al.

Assessment of the Images Visual Quality

The Figs. 5, 6, 7, 8, 9 and 10 provide the assessment of the visual quality of three images (Lena, Cameraman and Barbara) which are compressed by our approach.

(a) MSE=41.60,

(b)MSE=175.21,

(c)MSE=502.25,

rwPSNR=66.98,

rwPSNR=61.06,

rwPSNR=58.06,

mssim = 0,954

mssim= 0,785

mssim= 0,526

Fig. 5. The visual quality With wavelet transform of Lena

(a’)MSE=92.43,

(b’)MSE=175.75,

(c’)MSE=566.33,

rwPSNR=63.45,

rwPSNR=59.73,

rwPSNR=57.47,

mssim= 0.938

mssim= 0.753

mssim= 0.486

Fig. 6. The visual quality without wavelet transform of Lena

(a)MSE=93.58,

(b)MSE=135.972,

(c)MSE=403.24,

rwPSNR=61.47,

rwPSNR=59.20,

rwPSNR=57.72,

mssim= 0,928

mssim= 0.859

mssim= 0,664

Fig. 7. The visual quality with wavelet transform of Cameraman

Analysis of Image Compression Approaches Using Wavelet Transform

(a’)MSE=120.62,

(b’)MSE=211.84,

(c’)MSE=530.23,

rwPSNR=60.43,

rwPSNR=57.21,

rwPSNR=55.65,

mssim= 0.898

mssim= 0.739

mssim= 0.502

Fig. 8. The visual quality without wavelet transform of Cameraman

(a)MSE=246.09, rwPSNR=60.49, mssim=0,813

(b)MSE=258.69, rwPSNR=60.13, mssim=0,770

(c)MSE=469.14, rwPSNR=57.68, mssim=0,526

Fig. 9. The visual quality with wavelet transform of Barbara

(a’)MSE=265.42,

(b’)MSE=316.59,

rwPSNR= 57.56,

rwPSNR=56.61, mssim= 0.753

mssim= 0.792

(c’)MSE=507.3, rwPSNR=55.61, mssim= 0.541

Fig. 10. The visual quality without wavelet transform of Barbara

483

484

4.3

M. Rahali et al.

Comparison Between Our Approach and Various Methods

To evaluate our result, we use the standard image quality metrics for comparisons with other methods. The standard metrics is Peak Signal to Noise Ratio (PSNR) and Compression Ratio (CR). Tables 5 and 6 show the comparison between the proposed approach and various methods [10, 15, 17]. The methods were applied on standard images 256 * 256 (Lena and Cameraman): Neural Network one level, Neural Network two level [15], standard and modified Self-Organization Map SOM [10] and wavelet transform combined by Vector Quantization VQ [17]. Table 5. Comparison results for Lena image Methods Proposed method

Nbpp 1.39 1.8 Neural Network one level 2.32 Network neuron two levels 1.49 Modified SOM 1.84 Standard SOM 1.7 Wavelet transform and Vector Quantization 3.2

CR 82.59 78 71 81.3 77 79 60

PSNR 25.18 27.39 27.3 24.7 22.15 19.15 26.2

Table 6. Comparison results for Cameraman image Methods Nbpp Proposed method with wavelet transform 1.77 Modified SOM 2.32 Standard SOM 1.79

4.4

CR 77.8 71 77.6

PSNR 25.44 23.65 20.67

Simulation and Results

Our work is based on the comparison of different compression methods. The performance evaluation using various image quality metrics like PSNR, rwPSNR, MSE and MSSIM indicates that the best method is the wavelet transform and kohonen’s network. After comparing the two compression methods with and without wavelet transform, we can show the importance of using the wavelet transform in compression. In fact, the wavelet transform allows reducing the entropy of the image and separating its details to improve the quality of the reconstructed image according to the number of bits per pixel (Tables 3 and 4). The reconstructed image quality is acceptable; i.e. for rwPSNR it is more than 59 and for MSSIM it is between 0.7 and 1. The block effect is remarkable if the block size (BS) is higher than 256 and the degradation of the visual quality if the level of decomposition of the wavelet (J) is superior or equal to 2. For performance evaluation, the proposed method is compared with various methods [10, 15, 17]. The comparison is done using various measures such as PSNR, the number of bits per pixel Nbpp and compression ratio CR. From Tables 5 and 6, we can be deduced that the PSNR according of bits per pixel compression ratio of our

Analysis of Image Compression Approaches Using Wavelet Transform

485

method is better than of the some compression methods. From Table 5, the PSNR of our method is 25.18 and the CR is 82.59% but the PSNR of NNs with two levels [15] is 24.7 and CR is 81.3%. So, the PSNR and CR are better in proposed approach

5 Future Scope and Conclusion In this paper, we use an image compression approach based on the wavelet transform and the vector quantization by Kohonen’s network and the Huffman coding. We use two metrics to assess the reconstructed image quality: rwPSNR and MSSIM. We compare our method with other compression methods. The comparison is done using various standard measures such as PSNR and CR, it can be recognized that the PSNR according to CR of our method is better than some of the methods. To improve the reconstructed image quality and the compression ratio, we will add a phase of pre-treatment before the wavelet transform. We will use Weber-Fechner law and the A-law, which are used in the field of telephony, to quantify an image through the semi-logarithmic method.

References 1. Koff, D., Shulman, H.: An overview of digital compression of medical images: can we use lossy image compression in radiology? Can. Assoc. Radiol. J. 57(4), 211–217 (2006) 2. Zhang, N., Wu, X.: Lossless compression of color mosaic images. IEEE Trans. Image Process. 15(6), 1379–1388 (2006) 3. Kotyk, T., Ashour, A.S., Chakraborty, S., Balas, V.E.: Apoptosis analysis in classification paradigm: a neural network based approach. In: Healthy World Conference 2015 - A Healthy World for a Happy Life, At Kakinada (AP) India, September 2015 4. Kohonen, T.: The self-organizing map. Proc. IEEE 78(9), 1464–1480 (1990) 5. Mtimet, J., Benzarti, F., Amiri, H.: Lossless binary image coding using hybrid coding method. In: 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications, SETIT 2012, Sousse-Tunisia, 21–24 March 2012, pp. 448–452 (2012) 6. Pujar, J., Kadlaskar, L.: A new lossless method of image compression and decompression using Huffman coding techniques. J. Theor. Appl. Inf. Technol. 15(1/2), 18–23 (2010) 7. Ribés, J., Simon, B., Macq, B.: Combined Kohonen neural networks and discrete cosine transform method for iterated transformation theory. Signal Process.: Image Commun. 16(7), 643–656 (2001) 8. Hore, S., Bhattacharya, T., Dey, N., Hassanien, A.E., Banerjee, A., Chaudhuri, S.R.: A real time dactylology based feature extractrion for selective image encryption and artificial neural network. In: Image Feature Detectors and Descriptors. Studies in Computational Intelligence, vol. 630, pp. 203–226, February 2016 9. Raghuwanshi, P., Jain, A.: A review of image compression based on wavelet transform function and structure optimization technique. Int. J. Comput. Technol. Appl. 4, 527–532 (2013) 10. Boopathi, G.: An image compression approach using wavelet transform and modified self organizing map. Int. J. Comput. Appl. 28(2), 323–330 (2011)

486

M. Rahali et al.

11. Lewis, A.S., Knowles, G.: Image compression using the 2-D wavelet transform. IEEE Trans. Image Process. 1(2), 244–250 (1992) 12. Loukil, H., Kacem, M.H., Bouhlel, M.S.: A new image quality metric using system visual human characteristics. Int. J. Comput. Appl., 60(6) (2012) 13. Brooks, A.C., Zhao, X., Pappas, T.N.: Structural similarity quality metrics in a coding context: exploring the space of realistic distortions. IEEE Trans. Image Process. 17(8), 1261–1273 (2008) 14. Mathur, M.K., Loonker, S., Saxena, D.: Lossless Huffman coding technique for image compression and reconstruction using binary trees. Int. J. Comput. Technol. Appl. 3(1), 76–79 (2012) 15. Alshehri, S.A.: Neural network technique for image compression. IET Image Process., 1–5 (2015) 16. Zribi, M., Boujelbene, Y., Abdelkafi, I., Feki, R.: The “self-organizing maps of Kohonen in the medical classification”. In: 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications, SETIT 2012, Sousse-Tunisia, 21–24 March 2012, pp. 852–856 (2012) 17. Wang, H., Lu, L., Que, D., Luo, X.: Image compression based on wavelet transform and vector quantization. In: International Conference on Machine Learning and Cybernetics, Beijing, November 2002 18. Samanta, S., Ahmed, S.S., Salem, M.A., Chaudhuri, S.: Haralick features based automated glaucoma classification using back propagation neural network. In: International Conference on Frontiers of Intelligent Computing: Theory and Applications (FICTA), At Bhubaneswar, vol. 327, pp. 351–358 (2014) 19. Sudhakar, R., Karthiga, M.R., Jayaraman, S.: Image compression using coding of wavelet coefficients? A survey. ICGST-GVIP J. 5(6), 25–38 (2005) 20. Chaouch, D.E., Foitih, Z.A., Khelfi, M.F.: A sliding mode based control of 2DoF robot manipulator using neural network. In: 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications, SETIT 2012, Sousse-Tunisia, 21–24 March 2012, pp. 906-911 (2012) 21. Mili, F., Hamdi, M.: Comparative study of expansion functions for evolutionary hybrid functional link artificial neural networks for data mining and classification. Int. J. Hum. Mach. Interact. (IJHMI) 1(2), 44–56 (2014) 22. Dey, N., Mishra, G., Nandi, B., Pal, M., Das, A., Chaudhuri, S.: Wavelet based watermarked normal and abnormal heart sound identification using spectrogram analysis. In: Computational Intelligence & Computing Research (ICCIC), pp. 1–7, December 2012 23. Espartinez, A.S.: Self-fulfillment and participation of the human person in and through actions. Int. J. Hum. Mach. Interact. (IJHMI) 2(2), 28–31 (2015)

Application of Higham and Savitzky-Golay Filters to Nuclear Spectra Vansha Kher, Garvita Khajuria, Purnendu Prabhat(&), Meena Kohli, and Vinod K. Madan Model Institute of Engineering and Technology, Jammu, India [email protected], [email protected], [email protected], {purnendu.cse,meena.ece}@mietjammu.in

Abstract. A nuclear spectrum is a Type II digital signal. It is processed to extract the qualitative and the quantitative information about the radioisotopes. Noise reduction is usually a step in processing a spectrum, and smoothing filters are employed for this. Perhaps the most widely used filters for the spectral filter-ing since the last 50 years have been Savitzky-Golay (S-G) filters. Incidentally S-G filters are not well known to the digital signal processing community. In this paper the authors have employed rather unknown high fidelity Higham filters used by actuaries of the 19th century for filtering real observed nuclear spectra, and compared their performance with that of an optimal S-G filter. It was observed that the performance of the Higham filters and S-G filters compared favorably amongst each other. This paper describes theory of S-G and Higham filters, their application to nuclear spectra, and presents the results. Higham filters have potential for more applications. There is a scope of exploring more treasure of the past, and merging tools from various disciplines, for more interdisciplinary research. Keywords: Nuclear spectral processing
Type I and Type II signals filters
Savitzky-Golay filters
Higham filters




FIR

1 Introduction A nuclear spectrum e.g. a gamma-ray spectrum is processed to extract the qualitative and the quantitative information about the radioisotopes emitting gamma photons. In a typical observed nuclear spectrum over 90% of the information is lost thus greatly reducing the information capacity of a nuclear spectrometer. Processing of a spectrum usually involves employing a smoothing filter to reduce noise. It is desired that the filter should retain spectroscopic information content of the spectrum while attenuating noise. It is sometimes assumed that smoothing is manipulation of data without clear theoretical justification. Digital signal processing (DSP) gives a clear theoretical justification of a smoothing operation. In the frequency domain the signal containing the desired information gets concentrated in the low-frequency region while distortions extend up to p rad. The separation of signal and distortions appears most clear in the frequency domain. In any other domain like sequency domain, the clarity of separation gets obscured. A smoothing filter is thus adequately justified [1–8]. © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_42

488

V. Kher et al.

Perhaps the most widely used smoothing filters for the nuclear spectra since the last 50 years have been due to Savitzky and Golay [8]. These are linear phase, noncausal, FIR smoothing filters, also known as polynomial smoothing filters, or least squares smoothing filters. They have been widely used but not well known to the digital signal processing (DSP) community, and “only one out of about 20 had heard of them” [9]. One of the authors (VKM) interacted with both communities, DSP experts as well as nuclear scientists, and had similar experience, and was comforted to read lecture notes of Prof. Schafer [9], a highly respected DSP expert. Earlier the author happened to contribute to a bandwidth matching smoothing criterion for the S-G filters [3]. To bridge the gap between the two communities, he had contributed to a new classification of digital signals viz. Type I and Type II based on the fundamental problems of aliasing and quantization noise. The classification has widened the role of DSP to many disciplines [1, 4, 5]. In the larger perspective DSP techniques to nuclear spectral processing have hardly been used, though the benefits of DSP are widely accepted in diverse applications of science and technology. The least squares peak fitting functions entirely dominate the literature of this field, and these methods fall to a large extent within the category of “art-forms” rather than of “direct methods” [5]. Though DSP methods have not been widely used for the processing of nuclear spectra, they hold a good potential. Only a few researchers have reported the use of DSP techniques in this area. DSP methods for spectral processing, however, proved to be advantageous. They have enhanced the understanding of nuclear spectra while helping in the refinement, extension, and consolidation of various algorithms. The methods have further helped in developing new algorithms especially suited to the spectra and are extremely simple, easy, and fast [1–3, 5, 6]. In the 19th century FIR filters were employed for actuarial data processing [10–12]. The authors have explored two linear phase, high fidelity, FIR Higham filters from that era to filter real observed gamma spectra, and compared their performance with an optimal S-G filter for filtering the same spectra. The performance of the three filters was found favourable with respect to each other. To the authors knowledge Higham filters have neither been used in electrical engineering nor in nuclear spectrometry. Besides nuclear spectrometry, the filters hold a potential for their application to infrared spectrometry, analytical chemistry, heart rate monitoring using an accelerometer, speech and image processing etc. and that needs to be explored. This paper describes Type I and Type II signals, smoothing filters viz. S-G and Higham, and their application to filtering real observed nuclear spectra.

2 Type I and Type II Signals A digital signal can be represented as a mathematical function: yðxÞ ¼ A sin ð2p fx þ hÞ

ð1Þ

where both y(x) and its argument, x, are quantized. The quantity A is the amplitude, f is the frequency and h is the phase angle.

Application of Higham and Savitzky-Golay Filters to Nuclear Spectra

489

To widen the scope of DSP, a new classification of digital signals was given by Madan and Sinha [4, 5]. It is based on the fundamental problems of aliasing and quantization noise. Aliasing is a phenomenon where higher-frequency components take on the identity of lower-frequency components because of low sampling rate. It is addressed along the abscissa. Quantization error is the difference between the actual value of the signal and its quantized value. The noise power associated with the quantization of signals is called quantization noise. Type I digital signals are those where the problems of aliasing and quantization noise are addressed along the abscissa and the ordinate respectively. In Type II digital signals both the problems are addressed along the abscissa. DSP methods, widely used for processing Type I signals, are generally not used for processing Type II signals. The methods have, however, numerous advantages for processing Type II signals, e.g. nuclear spectra and age returns [4, 5]. The classification of signals arises because of the acquisition process of signals. While acquiring a DSP signals, both Type I and Type II, one has to ensure that aliasing should not be there, and quantization noise should be below a threshold. Once signals have been acquired, DSP techniques like filtering, deconvolution, Fourier domain processing etc. can be employed to process the signals. It may be mentioned that DSP techniques have demonstrated advantages, both theoretically and in implementation, for processing Type II signals like nuclear spectra and age returns. These methods, however, have hardly been employed. Figure 1 describes generation of a Type I digital signal like a speech signal, while Fig. 2 describes generation of a Type II digital signal like a nuclear spectrum. Figure 1 is well known to electrical engineers while Fig. 2 is not well known to the engineers.

analog signal in time domain

sampled – quantized signal Type I signal

sampled signal

x

ZOH

ADC

reconstructed signal in time domain DSP sampling pulses ZOH

Fig. 1. Acquisition and processing of a Type I signal

The Fig. 2 is described in more detail. Sampling instant is considered as the time when a given pulse arising out of nuclear radiation is detected, and immediately after the process of sampling, the pulse is quantized. Since the output of the nuclear analog to digital converter (NADC) will continue to be a time domain signal, as a matter of

490

V. Kher et al.

fact it is deliberately rotated 90° as the amplitude of each quantized pulse (or code, or energy channel in nuclear parlance) is quantified as the abscissa of histogram where all the pulses having assigned a particular code (or amplitude or energy) are counted in the same channel. It may be mentioned that a NADC has more stringent specification compared to those of commercial available ADC chips from manufacturers [1, 2]. preamplifier output in time domain histogram in the energy domain (Type II signal)

original spectrum

shaper output

ZOH

COUNTER

DSP

energy quantizer

DSP filtered spectrum ZOH

NADC 90° rotation: energy goes on the horizontal axis (time information lost)

Fig. 2. Acquisition and processing of a Type II signal

Let’s go through the signal theory for the two different situations. Figure 1 depicts electrical waveform in a common digital signal processing chain. Assume x(t) as the input waveform which is sampled at uniform intervals of time T. It is assumed that T is a constant irrespective of the variations in an analog signal, and due to jitter, which is assumed to be insignificant. The sampled signal xs(t) is given as: xs ð t Þ ¼

þ1 X

xðnT Þdðt  nT Þ

ð2Þ

n¼1

where, dðt  nT Þ is a unit impulse at time instant t = nT. In Fig. 1 quantization is done by a commercial ADC. In Fig. 2 it is done by a NADC, and the digital signals are condensed into a histogram thus generating a nuclear spectrum. The histogram is generated by employing the reconstruction filter. It is a zero order hold (ZOH) filter. It helps in maintaining the same level of the signal in a spectral channel.

Application of Higham and Savitzky-Golay Filters to Nuclear Spectra

491

ZOH is a low pass filter with a transfer function GT(s) which is given by: GT ðsÞ ¼

 1 1  esT s

ð3Þ

In Fig. 2, the second filter in the box DSP is a smoothing filter.

3 Gamma-Ray Spectrometer Figure 3 describes block diagram of a typical gamma-ray spectrometer. A gamma ray source emits gamma rays and they are detected by a HPGe detector. The charge collected by the detector, and then processed by the preamplifier produces sharp rising and pulses which decay with a large time constant. The pulses are shaped into Gaussian like pulses by the spectroscopy amplifier and they are condensed into histogram by a multichannel analyser (MCA) thus generating a nuclear spectrum. The spectrum is filtered and then further processing like background removal, deconvolution etc. is carried out to eventually extract the information about the radioisotoes emitting the nuclear radiation [1, 2].

HPGe Detector

Spectroscopy

Preamplifier

MCA

Spectral Processing

Amplifier

(including smoothing)

Source Detector Bias

Oscilloscope

Supply

Fig. 3. Block diagram of a gamma-ray spectrometer

4 Savitzky-Golay and Higham FIR Smoothing Filters In this section S-G and two Higham FIR filters are described. Both filters are noncausal, linear phase (symmetric) filters. S-G filters have been widely used and are easily found in the open literature albeit not in most DSP textbooks. Most DSP experts have not even heard of S-G filters [9]. As far Higham FIR filters, the authors have neither seen their application in electrical engineering nor in nuclear field. Described below is the theory of S-G and Higham filters. 4.1

Savitzky-Golay Filters

The S-G FIR smoothing filters, are also known as polynomial smoothing, or least-squares smoothing filters. It is a happy coincidence that the least squares polynomial fitting turns out to be FIR symmetric smoothing filters. S-G filters preserve high-frequency content representing peaks in a nuclear spectrum while attenuating noise.

492

V. Kher et al.

Savitzky and Golay put a polynomial to a set of input samples and then at the single point, evaluated the resulting polynomial with least squares fitting [8, 9]. Mathematically it proved to be a discrete convolution having fixed symmetric finite impulse response. The smoothing filter obtained by the least squares method helps in reduction of noise while preserving the shape and height of the nuclear spectral peaks. Consider for the moment the group of (2M + 1) samples of the signal x[n] centered at n = 0, we obtain the coefficients, ak , of a polynomial, pðnÞ, that minimizes pðnÞ ¼

XN k¼0

ak nk

ð4Þ

the mean-squared approximation error eN for the group of input samples centered on n = 0. eN ¼ ¼

XþM

ð pð nÞ  x ½ n Þ 2 !2 N X k ak n  x ½ n

n¼M þM X n¼M

ð5Þ

K¼0

The analysis is the same for any other group of (2M + 1) input samples. We will refer to M as the “half width” of the approximation interval. In S-G filter, the fitted polynomial to the sampled signal is similar to a fixed linear combination of the local set of input samples; the set of (2M + 1) input samples which are within the approximation interval are combined by a fixed set of weighting coefficients that can be calculated once for a given polynomial of order N, and remains same for another approximation interval of length 2M + 1. The polynomial coefficients, ak , represent impulse response h[n] of the FIR filter. Thus the output samples can be obtained by a discrete convolution method of the form: y ½ n ¼

þM X

h½mx½n  m

m¼M

¼

nX þM

ð6Þ h½n  mx½m

m¼nM

4.2

Higham Filters

Higham proposed filters that have high fidelity and are simple to implement. We briefly describe here Higham filters using the notation used by the then actuaries. Let n point boxcar average be represented as: 1 yi ¼ ½nxi n

ð7Þ

Application of Higham and Savitzky-Golay Filters to Nuclear Spectra

493

where xi is the input signal and yi is the output signal. The operator [n] is defined as: ½nxi ¼ xim þ xim þ 1 þ . . . þ xi þ m1 þ xi þ m

ð8Þ

Higham filters are depicted in Fig. 4. It has input signal xi and the output signal as yi. In operator form, it can be represented as: AT ¼ faA1 bA2 g

ð9Þ

where A1 and A2 are both summation filters, and a and b are scalars. Higham designed a number of filters including the following two. The authors could not find the detailed derivations of the Higham filters in the literature they had access to. The derivations were derived by the authors by hand using the Eqs. 10 and 11, and let these filters be called as Higham Filter 1 and Higham Filter 2 respectively. Both the filters resulted in 17 points smoothing filters [10–15]. 1 ½2½53 ½4½2  3½54 xi 125   64 4 4 2 ½5  ½5 ½9 xi yi ¼ 10000 300

ð10Þ

yi¼

A1

ð11Þ

a +

xi

yi A2

b

-

Fig. 4. Representation of a Higham filter

5 Experimental Results The real observed gamma-ray spectra were used in this paper. The spectra were taken from two earlier publications by one of the authors [3, 7]. These two spectral data sets were acquired using two different nuclear spectrometers. S-G, Higham Filter 1, and Higham Filter 2 filters were employed to filter the spectra. An optimal smoothing interval of 17 was chosen for all the smoothing filters [3]. It may be mentioned that in the earlier two publications [3, 7] employed only S-G filters to the spectra. The effect of smoothing on the spectra was investigated by employing all the three filters. The investigation included visual inspection, observing monotonicity in spectral peaks viz. it should be monotonically increasing before and monotonically decreasing after the peak position, change in full width at half maximum (FWHM), change in percentage peak areas using difference between sum of counts in the peak region of raw and smoothed spectra, and sum of the squared values of difference between the raw and

494

V. Kher et al.

smoothed spectral counts in the peak region, and it is called error squared values. The percentage change in peak area used the following formula: Percentage Change in Peak Area ¼ 100

Raw spectral area  smoothed spectral area raw spectral area

Fig. 5. Raw and smoothed spectra, shifted vertically arbitrarily, using Higham Filter 1 and S-G filter

Figure 5 depicts the raw, and the smoothed spectra obtained by employing the Higham Filter 1, and the S-G filter to one set of data, while Fig. 6 depicts the raw, and the smoothed spectra obtained by employing the Higham Filter 2, and the S-G filter to the other set of data. The visual inspection of the smoothed spectra employing Higham Filter 1, Higham Filter 2, and S-G on both the observed spectra indicated that smoothing effects were pronounced. The spectra were shifted vertically arbitrarily to get clarity of display in the figures. The channel number and counts in the figures correspond to energy and intensity respectively of nuclear radiation. From the printout of the filtered spectral data it was observed that the monotonicity as mentioned above was met in all the smoothed spectra. The changes in FWHM for all the smoothed spectral peaks were found by plotting the raw, and the smoothed spectra. It was observed that the change in FWHM was on the order of 0.1 channels for all the smoothed spectra. It is negligible. This indicates all the filters preserved high-frequency content representing peaks in a nuclear spectrum while attenuating noise. The Tables 1 and 2 depict sum of error squared values, and percentage change in peak areas respectively between the raw and the smoothed spectra. From the Table 1, it is observed that the sum of the error squared values for Higham filters were little better those for the S-G filters. It is observed from the Table 2 that the percentage change in peak areas by employing Higham and S-G filters were within 0.02% and hence negligible.

Application of Higham and Savitzky-Golay Filters to Nuclear Spectra

495

Fig. 6. Raw and smoothed spectra, shifted vertically arbitrarily, using Higham Filter 2 and S-G filter Table 1. Sum of error squared values Filter type Spectrum 1 Spectrum 2 Higham Filter 1 9474 39531 Higham Filter 2 9286 38639 S-G filter 11588 49805 Table 2. Percentage change in peak areas Filter type Spectrum 1 Higham Filter 1 −0.0158% Higham Filter 2 −0.0191% S-G filter −0.0126%

Spectrum 2 −0.0093% −0.0119% −0.0177%

However the relative error squared values depicted in Table 1, and the relative changes in percentage peak areas depicted in Table 2 can be considered as within the experimental limits. Hence from the above performance metrics it is concluded that the smoothing performance of all the three filters was favorable with respect to each other. Higham filters have potential for their application to many areas and that should be explored.

6 Conclusion For nuclear spectral filtering, perhaps the most widely used filters since the last 50 years have been S-G filters. However the filters are not well known to the DSP community, and “only one out of about 20 had heard of them” [7]. The authors have employed rather unknown Higham filters from the 19th century for filtering of real observed nuclear spectra. To the authors knowledge these filters

496

V. Kher et al.

have neither been employed in electrical engineering nor for nuclear applications. The performance of the Higham filters and S-G filter compared favorably with respect to each other. This paper demonstrated an interdisciplinary approach to nuclear spectral filtering. It has merged modern and 19th century tools from disciplines with encouraging application to nuclear spectrometry. Higham filters have potential for their application to infrared spectrometry, analytical chemistry, heart rate monitoring using an accelerometer, speech and image processing etc. and that needs to be explored. There is a scope of exploring more treasure of the past, and there is a need, perhaps more than ever, for merging of tools from various disciplines for the interdisciplinary research to be effective and fruitful. Acknowledgments. The authors would like to thank Director, MIET Prof. Ankur Gupta for his encouragement. This paper is respectfully dedicated to Prof. S.C. Dutta Roy.

References 1. Madan, V., Balas, M., Staderini, E.: Novel original digital signal processing techniques of nuclear spectra: a review. In: Kulczycki, P., Szafran, B. (eds.) Information Technology and Computational Physics (2016, designated for inclusion). Special Issue 2. Madan, V., Balas, M.: Nuclear spectrometry. Presented at the Congress of Information Technology, Computational and Experimental Physics (CITCEP 2015), Kraków, Poland, 18–20 December 2015, pp 194–198 (2015) 3. Kekre, H.B., Madan, V.K., Bairi, B.R.: A fourier transform method for the selection of a smoothing interval. Nucl. Instr. Meth. A279, 596–598 (1989) 4. Madan, V.K., Sinha, R.K.: Digital smoothing of census data employing fourier transforms. Comput. Stat. Data Anal. 20, 285–294 (1995) 5. Kekre, H.B., Madan, V.K.: Frequency domain and sequency domain filtering of nuclear spectral data. Nucl. Instr. Meth. A245, 542–546 (1986) 6. Madan, V.K., Abani, M.C., Bairi, B.R.: A digital processing method for analysis of complex gamma spectra. Nucl. Instr. Meth. A343, 616–622 (1994) 7. Madan, V.K., Sachdev, M.S.: A simpler method to reduce noise in uranium gamma spectra. Presented at the International Conference Radiation Safety in Uranium Mining, Saskatoon, Sask., Canada, 25–28 May (1992) 8. Savitzky, A., Golay, M.: Smoothing and differentiation of data by simplified least squares procedures. Anal. Chem. 36, 1627–1639 (1964) 9. Schafer, R.W.: What is a Savitzky Golay filter? IEEE Sig. Process. Mag. 28(34), 111–117 (2011) 10. Higham, J.A.: On the adjustment of mortality tables. J. Inst. Actuar. 23, 335–352 (1882) 11. Higham, J.A.: On the adjustment of mortality tables. J. Inst. Actuar. 24, 44–51 (1883) 12. Higham, J.A.: On the graduation of mortality tables. J. Inst. Actuar. 25, 15–24 (1884) 13. Finlaison, J.: Report on the Law of Mortality of the Government Life Annuitants. The House of Commons, London (1829) 14. Wilkinson, R.H.: Early history of the high-fidelity finite-impulse-response filter. In: IEE Proceedings, vol. 133, pt. G, no. 5, October 1986 15. Madan, V.K.: Digital signal processing: from antiquity to emerging paradigms with emphasis on nuclear spectrometry. ASET Colloquium, Tata Institute of Fundamental Research, Mumbai, 4 October 2013. http://www.tifr.res.in/*aset/

The Efficiency of Perceptual Quality Metrics 3D Meshes Based on the Human Visual System Nessrine Elloumi(&), Habiba Loukil Hadj Kacem, and Med Salim Bouhlel Research Unit: Sciences and Technologies of Image and Telecommunications, Higher Institute of Biotechnology, University of Sfax, Sfax, Tunisia [email protected], [email protected], [email protected]

Abstract. The representation of content as a 3D mesh is a very emerging technology. These three-dimensional meshes can be a scan of objects, characters or 3D scenes. Mesh quality is a determining factor in treatment of effectiveness, accuracy of results and rendering quality. You can show users these 3D meshes with a texture on the 3D mesh surface. The estimated quality by an observer is a very complex task related to the complexity of the Human Visual System (HVS). In this paper we present the efficiency of perceptual quality metrics 3D meshes based on the human visual system. Keywords: Perceptual quality


3D metric
3D meshes
HVS

1 Introduction For several years, researchers in image [1] processing and computer graphics have worked on the definition of an effective objective metric, able to predict the perceived quality of images, videos and 3D scenes [2]. To design such a system, it was necessary to study and determine the structures of visual content. This task is more complex when we look at 3D content represented in the form of triangular meshes. These 3D meshes are generally discrete packets of a non- uniform way to represent objects and scenes. Also, many operation can be performed on this 3D meshes [3], for example If we want to protect the 3D content we can use watermarking mesh algorithms for security purposes. It is also possible to apply a compression scheme to transmit on a low bandwidth network [4, 5] to reduce the size of 3D meshes [6]. Other operations on 3D meshes can be used, among which we mention: the quantification, enameling or deformation. These algorithms usually introduce inevitable distortion on the geometry sufaces mesh. It is therefore important to measure and evaluate the effect of these distortions that affect the quality triangular meshes.

2 The Perceptual Quality 3D Meshes The study of perceptual quality is an important task since most visual data are for a final human observer. Measuring the perceptual quality can be carried out using measurements made by observers (subjective measures) or measurements from © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_43

498

N. Elloumi et al.

algorithmic processes (objective measures). For each observer, quality can have a different definition of personal criteria. The main objective of measuring the perceptual quality is to analyze the behavior of the HVS [7] and quantify the quality as perceived by observers [8]. Human Vision System is based on two mechanisms: low-level mechanisms that affect the biophysical structure of the sensory system and high-level mechanisms related to human cognitive system.

3 Objective Measures of Quality Meshes The estimate of the perceptual quality is an important element in assessing the meshes quality after treatments. The objective of the perceptual quality metrics is to judge the quality of a mesh based on human perception characteristics [9]. To do this, there are different approaches the top-down approach considers the HVS as a black box and tries to imitate the behavior of the system from the perspective of inputs/outputs and the bottom-up approach based on simulation and imitation of each component of HVS. To develop a 3D metrics [10], researchers can use two alternatives: either they use existing metrics of perceptual quality 2D images [11–13] on two-dimensional projections of 3D meshes (metric-based images and videos) or they use develop metrics based-models that exploit the term mesh geometry and connectivity signal to evaluate the quality [14]. To estimate the quality of 3D meshes, other classification of the quality metrics is possible. * Metrics “Full – Reference”: The original model is available in full, in which case we will try to quantify the existing difference between the original image and the degraded version. * Metrics to “Reduced Reference”: The original image is available in its entirety, but is represented by a vector containing a reduced set of attributes. In this case, an attempt to quantify the difference existing between the vectors associated to the original image and the vector resulting from a distorted version. * Metrics “No – Reference”: The original image is not available. In this case, it is possible to work with a priori on the degradation types that may be encountered, or, conversely, by making no assumption about possible distortions. Many of these metrics exploit the proprieties of the human visual system. To compare two 2D images or 3D meshes, it is common to use the signal to noise ratio measurements or geometric distances. These geometrical measurements do not include the properties of human vision system. From the literature it is approved that the metric based on the HVS are more appropriate than others [15].

4 The Human Visual System The human visual system is a complex sensory system and not yet fully mastered. Nevertheless, it may be considered an information transcription system turned into usable data by the brain (Fig. 1). A conversion step can capture the received information and decoded into signals by the brain.

The Efficiency of Perceptual Quality Metrics 3D Meshes

499

Fig. 1. Human visual system: simplified

It is the role of the eye that converts light energy into sensory signals and then transmits them to the visual cortex via the optic nerves and geniculate nucleus. The visual cortex decrypts and processes the information. To better understand the flow of information and operation of the HVS, we recall the essential elements that are involved in the process of gathering and processing of visual information (Capture, conversion, transmission and processing of information). 4.1

Sensor

– The eye: Which an optical system whose primary role is to converge the light signals to the conversion zone and transmit the decoded form in the brain. It comprises several important elements including: – The cornea: is a convex spherical surface which separated the eye from the external environment. Its primary role is to focus the light rays received to the retina. – Iris: Adapts the light intensity by varying its opening. So acts as an optical diaphragm. – The crystalline: Allows to redirect the light flow to the retina where it’s arranged into photoreceptor cells (Cones and rods). It plays the role of variable focus lens. 4.2

Conversion

– The retina: It converts the light signal captured by the receiving photocells into electrical signals which are then transmitted to the visual cortex via the optic nerve. – The retina: is composed essentially of two types of receptors. – Cones: Which are responsive to detail. – Rods: Which are sensitive to low luminance (blurred vision and course) and intervene rather at night vision (monochrome). 4.3

Transmission

– The optic nerve: it transports information from the retina to the visual cortex, through the chiasm and lateral geniculation body.

500

N. Elloumi et al.

– The chiasm: the direction Information goes through from the right eye and the left eye, chiasm’s role is to transmit the information received to the lateral geniculation body. 4.4

Analysis and Decryption

– The visual cortex: His major role is the decrypting (or decoding) and the analyzing of the received signals.

5 The Human Visual System and the Perception of Visual Quality Most of the existing metrics follow the approach Top – Down to study the human visual system and imitate its behavior. The existing metrics has for objective to maximize the correlation of the results (profits) of prediction with the subjective scores. The subjective evaluation of the meshing quality is established by means of observers through psychometric experiments. Many databases were developed to have measures of the real structure of the 3D meshes perceptual quality. There are two important properties of the HVS that we have to consider in order to measure the perceptual quality of the 3D meshing [16]. 5.1

Contrast Sensitivity Function

The function of spatial contrast sensitivity (notes CSF) models the sensitivity of the human visual system to spatial frequencies of the visual stimulus. The first Research Campbell - Robson on CSF have shown that Works with the HVS selectivity on spatial frequencies, sensitivity achieved its Maximum Value around four cycles per degree of visual angle [17]. Figure 2(a) shows that the pixel intensity is modulated by a horizontal sine function. When the spatial frequency of the pixel increases in a logarithmical way (on the horizontal axis), the contrast increases from top to bottom. Although the change contrast is the same for all frequencies, we observe that the bars appear to be highest in the middle of the image following the shape of the sensitivity function in Fig. 2(b). This effect is not made by the image, it’s rather made by the property of frequency selectivity of the human vision system. In the context of dynamic visual content (2D videos, dynamic meshes, etc.), the CSF must be modulated by stimuli speed. This is due to the variability of the contrast sensitivity depending on the speed of the stimulus movement. Since 1977, Kelly has presented an experimental study for modulating the function of contrast sensitivity affected by the test conditions (see Fig. 3(a)). In 1998, Daly has developed a dynamic model of CSF through a time function of CSF tends to move to lower frequencies. These studies were performed for achromatic data. Other studies have integrated color effect on the CSF: The contrast sensitivity (Fig. 3(b)). With increasing speed, the CSF

501

Contrast

The Efficiency of Perceptual Quality Metrics 3D Meshes

Fréquency SpaƟal

(a)

(b) Fig. 2. The spectral properties of the HVS: (A) - graph illustrating Campbell- Robson contrast sensitivity (CSF), (b) - curve of HVS sensitivity as a function of spatial frequency.

curve tends to shift towards low frequencies. These studies were realized to achromatic data [18]. 5.2

The Masking Effect

Masking is the effect of modification of the component bordures visibility in a multimedia content (masked signal) by the presence of another component (masking signal). The magnitude of this effect is measured by the change of the masked signal visibility with or without the presence of the masking signal [19]. Masking may intervene in the same frequency band (in-band masking) or between different frequency bands (inter-band masking). This type of mask in literature, called the entropy masking effect [20]. There is another type of contrast masking named masking contrast connected to the

502

N. Elloumi et al.

Fig. 3. Kelly Illustration (a) - for the presentation of the combination of the effect of spatial and temporal frequency on contrast sensitivity, and (b) - the CSF illustration depending on the speed. The time function of CSF in (b) is calculated from an empirical equation. The speeds V are measured in degree/second.

change of the surface visibility as a function of the contrast values. The spatial masking effect is often linked to the concept of surface roughness for 3D meshes [21]. 5.3

The Perceptual Quality 3D Mesh

To measure the distance between two 3D meshes [22] many 3D metrics exist. We represent in the following sheet, the perceptual metrics developed for 3D meshes [23]. 5.3.1 The 3D Watermarking Perception Metric Corsini et al. have developed a new metric of quality [24]. Their approach, called 3DWPM (3D Watermarking Perception Metric), is based on the calculation of the perceptual distance between two meshes resting on surface roughness. This measured 3DWPM distance between two meshes M1 and M2 are identified by:

The Efficiency of Perceptual Quality Metrics 3D Meshes

3DWPMðM1; M2Þ ¼ logð

qðM2  qðM1Þ þ KÞ  logðkÞ qðM1Þ

503

ð1Þ

where qðM1 Þ and qðM2 Þ measure the overall roughness of the two meshes and k is a constant varied digital. Two variants of 3DWPM were developed using descriptors of roughness differences. The first descriptor of roughness is used to 3DWPM1, is inspired by that of Lavoué [25]. The roughness measurement is calculated through measuring the dihedral angles between the normal facets in vicinity. The normal facet of a smooth surface does not vary greatly. In contrast, on textured regions (rough) these normal vary more significantly. A Multi - scale analysis of these entities is considered [26] to evaluate dihedral angles using the direct vicinity (1 ring) and the extended vicinity (1 ring, 2 rings, etc.). The second roughness measurement adopted by Corsini et al. is 3DWPM2, it is based on estimating of the roughness of surfaces by Wang et al. [27]. Their approach is based on the comparison of a mesh and smoothed versions of the same mesh. Smooth regions correspond to small differences while the rough areas have more significant differences. 5.3.2 The Mesh Structural Distortion Measure Metric The MSDM (Mesh Structural Distortion Measure) uses the amplitude of the average curvature of the 3D mesh surface to quantify the perceptual distortion. This metric has been improved recently under the name MSDM2 by integrating multi-scale analyses [25]. In 2006, Lavoué et al. have introduced a structured distortion measurement called Structural Mesh Distortion Measure (MSDM) [28]. This quality measure was inspired by the quality measurement of 2D pictures SSIM (Structural SIMilarity index) introduced by Corsini et al. [24]. The MSDM is based on the statistical difference of the average amplitude curves to measure the perceptual difference of two meshes. The average curvatures (CMsi) are calculated for each as the average of the minimum and maximum curvatures: CMsi ¼

ðCMsi Þ ¼ jCmin; sij þ jCmax; sij 2

ð2Þ

The amplitude of minimum and maximum curvatures (jCmin; sij and jCmax; sij) are deducted from the fair values of the curvature tensor. The approximation of the curvature tensor used, is that introduced by Cohen-Steiner et al. on spatial vicinity defined by the geodesic disc resulting from the projection of a sphere of radius h on the surface of mesh. The average and standard deviation of the average curvature in a spatial window w containing n peaks, denoted respectively lw and rw are defined as: 1X CMsi; si2w n rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1X rw ¼ ðCMsi  lw Þ2 ; si2w n lw ¼

ð3Þ ð4Þ

504

N. Elloumi et al.

The covariance between the curvatures of two window s w1 and w2 of the compared meshes M1 and M2 is defined as: r w1 w 2 ¼

rww11 w2 þ rww21 w2 2

ð5Þ

1 where rw w1 w2 defines the covariance calculated on the window w1:

rww11 w2 ¼

     1X w1 w2  l CM  l CM w w S S2i 1 2 1i S1i 2w1 n

ð6Þ

where S2i is the nearest summit to S1i on S1i the covariance rw2 w1w2 on the window w2 is calculated in the same way. To calculate the overall distance between two meshes, MSDM based on local distances, we note MSDML: 1

MSDML ¼ ða  LðsÞa þ b  CðsÞa þ c  M ðsÞa Þa

ð7Þ

With a, b, c and a are scalars for combining different quantities. The parameters L, C, M represent respectively the differences bends, contrasts and structures measured by: LðsÞ ¼

jjlw1  lw2 jj MAX ðlw1 ; lw2 Þ jjrw1  rw2 jj MAXðrw1 ; rw2 Þ

ð9Þ

jjrw1 rw2  rw1 rw2 jj rw1 rw2

ð10Þ

Cðw1; w2Þ ¼ M ðw1; w2Þ ¼

ð8Þ

where w represents the number of local windows on mesh surfaces to compare, and a parameter that has the same value as in the Eq. 7. An improved version of MSDM named MSDM2 was proposed in 2011 by integrating a multi-scale analysis [29]. MSDM2 also compares two meshes that do not share the same connectivity through a matching step through the structure data AABB tree implemented in the CGAL [30]. We note that to establish a measure of perceptual symmetrical distance, MSDM2 calculates the average of the two distances unidirectional (respectively M1 to M2 and M2 to M1 ). MSDM and MSDM2 are based only on statistics of the curvatures magnitudes. 5.3.3 The Fast Mesh Perceptual Distance Metric In 2012, a new metric for estimating the perceptual quality was introduced par Corsini et al. [31]. It is entitled Fast Mesh Perceptual Distance (FMPD). The type of this metric is reduced-reference, it is based on the comparison of overall roughness measurements calculated on the two grids to compare. The roughness descriptor retained in FMPD is

The Efficiency of Perceptual Quality Metrics 3D Meshes

505

derived from Laplacian of the Gaussian curvature. For each summit Si the Gaussian curvature (noted CGi) is established by the following equation:   X   CGi ¼ 2p  a ð11Þ  j jN ðF Þ i

or Ni (F) represents all vicinity facets at the summit Si and the j angle aj represent the crossing facet with the actual summit. The local roughness, calculated on a top si is approximated by the quantity: P    j2Ni ðSÞ Dij CGj   RLi ¼ CGi þ  D

ð12Þ

ii

where Ni ðsÞ is the set of neighboring vertices aware summit, the matrix D it’s the array of discrete Laplace established by:     cot bij þ cot bij pourjNi ðsÞ Dij ¼ 2 X Dii ¼  D j ij

ð13Þ ð14Þ

where bij and bij′ are the opposite angles in which the peak relies si to sj. Figures 4(b) and (d) represent respectively the color card of the roughness measurement on the surface of the Armadillo and Venus meshes. The warm colors match at high RL values, cold colors correspond to low values of RL. We note that this descriptor is able to properly evaluate and classify the surfaces of meshes into smooth and rough areas regions. FMPD in the local roughness measurements are modulated by a power function [32] to establish local values of roughness RLFi in each peak si. This modulation is performed to capture the spatial masking effect on the mesh surfaces.

Fig. 4. (a) - and (c) - respectively present the 3D meshes of Armadillo and Venus, (b) - and (d) present colors of the cards from the respective surfaces roughness Armadillo of mesh and mesh Venus.

506

N. Elloumi et al.

The overall roughness is subsequently calculated by a steady sum of the local measurements: P i RLFi ai RG ¼ P ð15Þ i ai where ai is a coefficient connected to the area of the incident facets at the summit si the measure of the perceptual distance FMPD is established as the difference between the values of global roughness RG1 and RG2 of two meshes M1 and M2 to compare: FMPD ¼ cjRG1  RG2 j

ð16Þ

With c as a scalar set a scale parameter for the FMPD in the interval distances [0; 1]. 5.3.4 The Dihedral Angle Mesh Error Metric Recently, Vasa and Rus have developed a new metric of perceptual quality for static 3D mesh based on the dihedral angles [33]. This metric, named DAME (Dihedral Angle Mesh Error) offers a compromise between the complexity and the prediction efficiency. On each pair of triangles (t1 = {S1S2S3}; t2 = {S3S2S4}) neighbors of the mesh, the dihedral angle oriented are calculated use this expression: Dt1t2¼arccosðn1;n2Þsgnðn1:ðs4s3ÞÞ

ð17Þ

where n1 and n2 represent respectively the surface normal associated with t1 and t2 triangles. Measurement of DAME perceptual distance is calculated as follows: 1 jjXjj

ð18Þ

  Dt1t2  Dt2t2 
mt1t2 :ðw1 þ w2 Þ

ð19Þ

DAME ¼ X ft1;t2g2X

where Dt1t2 and Dt1t2 respectively represent the measurements of the dihedral angles of the two meshes to compare, m and w represent respectively a masking modeling function and a visibility coefficient.

6 Conclusion In this paper, we have studied the importance of perceptual quality 3D meshes compared to the objective measures meshes. We have also enumerated the different existing approaches which are based on the metrics of the perceptual quality 3D meshes. We have also detailed various properties of the HVS, such as the CSF and the masking effect. It is important to consider these properties for guiding quality measures development process perceptual 3D meshes. These existing perceptual objective

The Efficiency of Perceptual Quality Metrics 3D Meshes

507

metrics mainly considers the spatial masking effect to establish quality metrics collected for 3D meshes. Finally we have itemized metrics which are based on the HVS properties.

References 1. Lecuire, V., Duran-Faundez, C., Krommenacker, N.: Energy-efficient image transmission in sensor networks. Int. J. Sens. Netw. 4(1/2), 37–47 (2008) 2. Salehpour, M., Behrad, A.: 3D face reconstruction by KLT feature extraction and model consistency match refining and growing. In: 2012 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications (SETIT), Sousse, pp. 297–302 (2012) 3. Chan, A.T., Gamino, A.: Integration of assistive technologies into 3D simulations: an exploratory study. In: Information Technology: New Generations, Advances in Intelligent Systems and Computing, vol. 448, pp. 425–437. Springer (2016). ISBN 978-3-319-18295-7 4. El-Bendary, M.A.M., El-Tokhy, M., Kazemian, H.B.: Efficient image transmission over low-power IEEE802.15.1 network over correlated fading channels. In: The 6th International Conferences: Sciences of Electronics, Technologies of Information and Telecommunications “SETIT 2012”, Mars 2012, Sousse-Tunisie, IEEE Conferences, pp. 563–567 (2012). doi:10. 1109/SETIT.2012.6481973 5. Abderrahim, Z., Techini, E., Bouhlel, M.S.: State of the art: compression of 3D meshes. Int. J. Comput. Trends Technol. (IJCTT) 4(6), 765–770 (2012) 6. Tang, H., Joshi, N., Kapoor, A.: Learning a blind measure of perceptual image quality. In: Proceedings of the 2011 IEEE Conference on Computer Vision and Pattern Recognition, pp. 305–312 (2011) 7. Hemanth, D.J., Balas, V.E., Anitha, J.: Hybrid neuro-fuzzy approaches for abnormality detection in retinal images. In: Proceedings of the 6th International Workshop Soft Computing Applications, SOFA 2014, Timisoara, Romania, 24–26 July 2014, pp. 295–305 (2014) 8. El-Bendary, M.A.M., El-Tokhy, M., Shawki, F., Abd-El-Samie, F.E.: Studying the throughput efficiency of JPEG image transmission over mobile IEEE 802.15.1 network using EDR packets. In: The 6th International Conferences: Sciences of Electronics, Technologies of Information and Telecommunications “SETIT 2012”, Mars 2012, Sousse-Tunisie, IEEE Conferences, pp. 573–577 (2012). doi:10.1109/SETIT.2012. 6481975 9. Escribano-Barreno, J., García-Muñoz, J.: Integrated metrics handling in open source software quality management platforms. In: Information Technology: New Generations. Advances in Intelligent Systems and Computing, vol. 448, pp. 509–518. Springer (2016). ISBN 978-3-319-18295-7 10. Triki, N., Kallel, M., Bouhlel, M.S.: Imaging and HMI, fondations and complementarities. In: The 6th International Conferences: Sciences of Electronics, Technologies of Information and Telecommunications, SETIT 2012, Mars 2012, Sousse-Tunisie, IEEE Conferences, pp. 25–29 (2012). doi:10.1109/SETIT.2012.6481884 11. Nandi, D., Ashour, A.S., Samanta, S., Chakraborty, S., Salem, M.A., Dey, N.: Principal component analysis in medical image processing: a study. Int. J. Image Min. 1(1), 65–86 (2015) 12. Cho, J.-W., Prost, R., Jung, H.-Y.: An oblivious watermarking for 3-D polygonal meshes using distribution of vertex norms. IEEE Trans. Sig. Process. 55(1), 142–155 (2007)

508

N. Elloumi et al.

13. Wang, K., Lavoué, G., Denis, F., Baskurt, A.: Robustand blind mesh watermarking based on volume moments. Comput. Graph. 35(1), 1–19 (2011) 14. Abderrahim, Z., Techini, E., Bouhlel, M.S.: Progressive compression of 3D objects with an adaptive quantization. Int. J. Comput. Sci. Issues (IJCSI) 10(2), 504–511 (2013) 15. Campbell, F.-W., Robson, J.-G.: Application of Fourier analysis to the visibility of gratings. J. Physiol. 197, 551–566 (1968) 16. Hirai, K., Tsumura, N., Nakaguchi, T., Miyake, Y., Tominaga, S.: Spatio-velocity contrast sensitivity functions and video quality assessment. In: International Symposium on Intelligent Signal Processing and Communication Systems, pp. 1–4 (2010) 17. Ninassi, A., Meur, O.L., Le Callet, P., Barba, D.: On the performance of human visual system based image quality assessment metric using wavelet Domain. In: Proceedings of the SPIE Human Vision and Electronic Imaging (2008) 18. Ninassi, A., Meur, O.L., Le Callet, P., Barba, D.: Which semi-local visual masking model for wavelet based image quality metric In: Proceedings of IEEE International Conference on Image Processing, pp. 1180–1183 (2008) 19. Fernandez-Maloigne, C., Larabi, M.-C., Bringier, B., Richard, N.: Spatio temporal characteristics of the human color perception for digital quality assessment. In: Proceedings of International Symposium on Signals, Circuits and Systems, pp. 203–206 (2005) 20. Kelly, D.-H.: Visual contrast sensitivity. Opt. Acta: Int. J. Opt. 24(2), 107–129 (1977) 21. Daly, S.-J.: Engineering observations from spatio velocity and spatio temporal visual models. In: Proceedings of the SPIE Human Vision and Electronic Imaging III, vol. 3299, pp. 180–191 (1998) 22. Wang, Z., Bovik, A.-C.: Modern Image Quality Assessment. Morgan & Claypool, San Rafael (2006) 23. Rogowitz, B.-E., Rushmeier, H.-E.: Are image quality metrics adequate to evaluate the quality of geometric objects. In: Proceedings of SPIE Human Vision and Electronic Imaging, pp. 340–348 (2001) 24. Corsini, M., Drelie Gelasca, E., Ebrahimi, T., Barni, M.: Watermarked 3-D mesh quality assessment. IEEE Trans. Multimedia 9(2), 247–256 (2007) 25. Lavoué, G.: A multiscale metric for 3D mesh visual quality assessment. Comput. Graph. Forum 30(5), 1427–1437 (2011) 26. Lavoué, G., Drelie Gelasca, E., Dupont, F., Baskurt, A., Ebrahimi, T.: Perceptually driven 3D distance metrics with application to watermarking. In: Proceedings of SPIE Electronic Imaging (2006) 27. Wang, Z., Bovik, A.-C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process. 13(4), 600–612 (2004) 28. Alliez, P., Tayeb, S., Wormser, C.: 3D fast intersection and distance computation (AABB tree). In: CGAL User and Reference Manual (2012) 29. Wu, J.-H., Hu, S.-M., Tai, C.-L., Sun, J.-G.: An effective feature-preserving mesh simplification scheme based on face constriction. In: Pacific Conference on Computer Graphics and Applications, pp. 12–21 (2001) 30. Gelasca, E.-D., Ebrahimi, T.: Objective evaluation of the perceptual quality of 3D watermarking. In: IEEE International Conference on Image Processing, pp. 241–244 (2005) 31. Corsini, M., Gelasca, E.-D., Ebrahimi, T.: A multi-scale roughness metric for 3D watermarking quality assessment. In: Workshop on Image Analysis for Multimedia Interactive Services (2005) 32. Wang, K., Torkhani, F., Montanvert, A.: A fast roughness-based approach to the assessment of 3D mesh visual quality. Comput. Graph. 36(7), 808–818 (2012) 33. Vása, L., Rus, J.: Dihedral angle mesh error: a fast perception correlated distortion measure for fixed connectivity triangle meshes. Comput. Graph. Forum 31(5), 1715–1724 (2012)

Fireworks Algorithm Based Image Registration Silviu-Ioan Bejinariu1(&), Hariton Costin1,2, Florin Rotaru1, Ramona Luca1, Cristina Diana Niţă1, and Camelia Lazăr1 1

Institute of Computer Science, Romanian Academy Iaşi Branch, Iasi, Romania {silviu.bejinariu,florin.rotaru, ramona.luca,cristina.nita, camelia.lazar}@iit.academiaromana-is.ro, [email protected] 2 Faculty of Medical Bioengineering, “Grigore T. Popa” University of Medicine and Pharmacy, Iasi, Romania

Abstract. In the Image Processing (IP) domain, optimization algorithms have to be applied in many cases. Nature-inspired heuristics allow obtaining near optimal solutions using lower computing resources. In this paper the Fireworks Algorithm (FWA) behavior is studied for Image Registration (IR) problems. The IR results accuracy is analyzed for different types of images, mainly in case of pixel based registration using the Normalized Mutual Information. FWA is compared to Particle Swarming (PSO), Cuckoo Search (CSA) and Genetic Algorithms (GA) in terms of results accuracy and number of objective function evaluations required to obtain the optimal geometric transform parameters. Because the pixel based IR may fail in case of images containing graphic drawings, a features based IR approach is proposed for this class of images. Comparing to other nature inspired algorithms, FWA performances are close to those of PSO and CSA in terms of accuracy. Considering the required computing time, that is determined by the number of cost function evaluations, FWA is little slower than PSO and much faster than CSA and GA. Keywords: Fireworks Algorithm


Optimization
Image registration

1 Introduction Image registration (IR) is the process of geometric overlaying or alignment of two or more images of the same scene for subsequent common processing [1]. The goal of an IR procedure is to approximate the parameters of the geometric transform to be applied to a source image such that it is aligned to a model image. The overlaying quality is evaluated using a similarity criterion which must be maximized using an optimization algorithm. In case of a large number of parameters, or of complex similarity measures, the optimal parameters may be difficult to obtain. Based on the strategies of live beings to survive, find food or avoid obstacles, or modeling other natural phenomena, the nature-inspired metaheuristics offer the possibility to obtain faster a near optimal solution [2]. © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_44

510

S.-I. Bejinariu et al.

A relatively new nature inspired algorithm is the Fireworks Algorithm (FWA) proposed in [3] and deeply analyzed in [4]. A detailed study and some enhancements of the operators (strategies) used in FWA are presented in [5]. Other versions of FWA are: an adaptive version of FWA proposed in [6] and a cooperative strategy among fireworks in [7]. Multi-objective versions of FWA were also developed [8]. In most papers the algorithm’s performances are studied using benchmark functions. The IR issue was addressed in [9–11] using Bacterial Foraging Optimization and in [12] using Bat and Cuckoo Search algorithms. In [13] the behavior of Bacterial Foraging Optimization, Particle Swarming, Multi-swarming and Firefly algorithms in multi-threshold image segmentation was studied. In [14] single and multi-objective FWA convergence was analyzed using test functions. The paper is organized as follows. In the second section FWA is briefly described. The third section describes the IR problem and the experiments made using different types of images, including an analysis of an IR failure case. In the fourth section, the accuracy of FWA results is compared to that of the results obtained using the Particle Swarming, Cuckoo Search and Genetic Algorithms as optimization procedure. In the fifth section, a features based IR approach is shortly described. It allows the registration of images that contain graphic drawings, that fails in case of pixel based registration. The last section concludes the paper.

2 Fireworks Algorithm The Fireworks Algorithm (FWA) belongs to the Swarm Intelligence algorithms category and it is inspired from the fireworks explosion simulation [3, 4]. As in other swarming algorithms, the possible solutions are encoded as position of individuals that evolve in the problem domain. The evolution is simulated by selection of sparks resulted in the fireworks explosion process. The objective to optimize is evaluated in all the reached positions and the best obtained position is considered as being the problem solution [14]. FWA was initially developed for single-objective optimization problems, in which the parameter value that minimize (or maximize) a fitness function have to be determined: minx2S f ð xÞ

ð1Þ

where S is the problem domain (constraints). FWA is an iterative algorithm which starts with a set of individuals (fireworks) randomly placed in the problem domain. For each iteration (epoch) of FWA, a variable number of descendants (sparks) is generated for each firework. The descendants are evaluated and some of them became fireworks in the next evolution step
(epoch).  Let’s consider X ¼ fxi ; i ¼ 1; . . .N g a set of N fireworks, and xi ¼ x1i ; . . .; xdi their positions in the problem domain, where d is the dimension of the problem.

Fireworks Algorithm Based Image Registration

511

The general structure of the algorithm is described below [4]:

Each step of the FWA loop involves strategies that are important for the algorithm’s performances. The fireworks explosion which ensures the quality of the next fireworks generation is characterized by strength, amplitude and displacement. The explosion strength refers to the number of sparks Si of each firework. It depends on the current firework quality, and it is computed as: fworst  f ðxi Þ þ e Si ¼ m  Pn 1 ðfworst  f ðxi ÞÞ þ e

ð2Þ

where m is the total number of sparks, fworst is the worst fitness value of the current fireworks set, f ðxi Þ is the fitness value of the ith firework and e is a constant used to avoid zero-division operations. To keep the number of sparks in a reasonable interval, minimum and maximum values that depend on m are imposed for Si . The amplitude Ai of the ith firework sparks is used to avoid the convergence to local solutions. The amplitude is higher as the distance to the best firework increases and it is computed as: f ðxi Þ  fbest þ e Ai ¼ A  Pn 1 ðf ðxi Þ  fbest Þ þ e

ð3Þ

where A is the desired sum of amplitudes, fbest is the best fitness value in the current fireworks set. Two types of sparks are created in the ith firework explosion: regular and Gaussian. Regular sparks are obtained from the parent firework by adding a uniform random number in the interval ðAi ; Ai Þ to some randomly chosen coordinates. xki ¼ xki þ U

ð4Þ

where k are the randomly chosen coordinates and U is the uniform random value.

512

S.-I. Bejinariu et al.

The Gaussian sparks assure the diversity of the next fireworks generation by altering some randomly chosen coordinates of some randomly chosen fireworks by random numbers with Gaussian distribution: xki ¼ xki  g

ð5Þ

where k are the randomly chosen coordinates and g are random numbers. If the generated sparks are placed outside the problem domain they must be re-positioned inside. The fitness function is evaluated in the all new sparks and the next generation fireworks population is selected. The best, the worst and the other N  2 randomly chosen fireworks replace the population of fireworks.

3 FWA Based Image Registration Image registration is the process of geometric overlaying or alignment of two or more images of the same scene taken at different times, from different viewpoints, and/or by different sensors [1]. There are two different approaches of IR: pixel level and features level registration. In both cases, IR is an optimization procedure because it requires to approximate the parameters of a geometric transform which applied to a source image minimize a similarity measure computed between the transformed source image and a model image. In the FWA based IR the possible solutions are encoded as positions of the fireworks and the objective to optimize is the Normalized Mutual Information (NMI). The fireworks evolve in a multidimensional space whose dimension is equal to the number of parameters in the geometric transform to be approximated. In the next paragraphs the Normalized Cross-Correlation (NCC) is also used to evaluate the results, but it is not used as optimization criteria. The FWA based IR procedure was evaluated using the images available in the ‘Miscellaneous’ volume of USC-SIPI database [15]. All used images were resized to 256  256 resolution, and color depth was reduced to 256 gray levels if not already so. In the following paragraphs the results obtained using four model images from this database are presented. Two of these images contain photos (Fig. 1a and b), one contains graphics (Fig. 1c) and the last one is mixed (Fig. 1d).

a. ‘Lena’ model image b. ‘Peppers’ model image

c. ‘Ruler’ model image

Fig. 1. Original test images, from [15]

d. ‘Testpat’ model image

Fireworks Algorithm Based Image Registration

513

The source images were artificially generated by applying to model images an affine transform with the following parameters: rotation by h ¼ 10 against the rotation center cx ¼ 30; cy ¼ 30 and scaling by the factor s ¼ 0:8. The computed inverse affine transform has the following parameters: rotation by h0 ¼ 10 against the same center and scaling factor s0 ¼ 1:250. FWA was applied to register the source images (Fig. 2) to model images (Fig. 1). The parameters of the algorithm were chosen as follows: – – – – – –

fitness function: NMI, number of epochs = 250, number of fireworks = 5, maximum amplitude = 40, regular sparks factor = 10 and number of Gaussian sparks = 5.

These values of FWA parameters were chosen accordingly to those proposed in [4] and they were validated by our experiments. Because optimal value of NMI is unknown, there was no stop criterion other than the number of epochs. It must be noticed that the fitness function evaluation requires to apply the approximated geometric transform that correspond to the firework position and then to compute the similarity measure between the model image and transformed source image. This evaluation is time consuming; in fact more than 90% of IR execution time is spent in the fitness function evaluation [10].

a. ‘Lena’ source image b. ‘Peppers’ source image

c. ‘Ruler’ source image

d. ‘Testpat’ source image

Fig. 2. Source images obtained by applying the geometric transform to model images

Even if the number of individuals (fireworks) and iteration (epochs) are reduced comparing to other nature-inspired IR procedures, the results are quite good. The final results evaluation is performed using two different similarity measures: NMI and NCC. The columns named Expected contain the NMI/NCC values computed between the model image and source image transformed using the computed inverse transform. The columns named Approximated contain the NMI/NCC values computed between the model image and source image transformed using the approximated inverse transform. The column ‘#evals’ contains the total number of cost function evaluations, column ‘Best eval’ the number of the evaluation in which the best fitness was reached and the last column contains the execution time in seconds – not so relevant while it depends on computer’s performances.

514

S.-I. Bejinariu et al.

By analyzing the values in Tables 1 and 2, the following conclusions can be drawn: – In most cases the approximated value of the similarity measure is greater than the expected value. This can be explained by the two geometric transforms applied to the image which means that pixels values were interpolated two times. – The parameters of the approximated inverse transform (Table 2) are very close to those of the computed inverse. – In case of ‘Ruler’ image the registration process failed. The cause is discussed below. – The best value was obtained enough late in the iterative process, which means that better solutions can be obtained by increasing the number of iterations. Regarding the number of epochs and number of fireworks used in FWA it must be said that a reduced number of fireworks (3 or 4) requires a considerably increase of the number of epochs in order to obtain similar results. By increasing the number of epochs and keeping constant the number of fireworks (=5), the results are improved but not significantly while the processing time greatly increases. – The application was tested on an Intel Core i3, 2.4 GHz, 6 GB RAM based computer. The execution time is presented for the case of parallel execution. Even if not present in the Table 1, the sequential execution time is about 30% greater with the observation that other applications were running in the same time.

Table 1. Results evaluation of IR process Image

NMI Expected Lena 1.3252 Peppers 1.3463 Ruler 1.1147 Testpat 1.5014

Approximated 1.3253 1.3466 1.0492 1.5014

NCC Expected 0.9765 0.9675 0.9699 0.9525

# evals Best eval Time (sec) Approximated 0.9766 0.9675 0.3511 0.9525

14369 14390 14109 14349

14062 14329 13867 14223

27.300 27.986 26.567 29.422

Table 2. Parameters of the approximated inverse transform h0

Image

c0x

c0y

s0

Lena Peppers Ruler Testpat

30.1959 30.0360 40.7941 29.9996

30.0245 −9.9944 1.2504 30.0073 −10.0032 1.2500 48.6515 −9.9937 1.2519 30.0161 −10.0012 1.2500

The registered images are depicted in Fig. 3. The failure of ‘Ruler’ image (which contains graphics) registration is also visible (Fig. 3c).

Fireworks Algorithm Based Image Registration

a. ‘Lena’ registered image

b. ‘Peppers’ registered c. ‘Ruler’ registered image image

515

d. ‘Testpat’ registered image

Fig. 3. Registered images obtained by applying the approximated inverse transform

3.1

IR Failure Analysis

As it was seen above, the IR of ‘Ruler’ image failed. Let’s analyze the histograms of the four images (Fig. 4). Only 5 levels of gray are used in the ‘Ruler’ image, most of them corresponding to the image background. By applying the affine transform to the model, most pixels have the same or close values to those in the original, leading to small variations of the similarity measure. This is visible in Fig. 5 which contains for all images the variation of NMI depending on the rotation angle and scale factor, keeping the rotation center fixed and equal to the values determined in the computed inverse transform. The peak of NMI surface is visible for all images, but the variation is steep close to the maximum value in case of ‘Ruler’ while the growth is smoother for the other images. It must be specified also that these surfaces have about 400 local maximum values in case of ‘Lena’ and ‘Peppers’, 122 local maximum in case of ‘Testpat’ and about 1040 maximum values in case of ‘Ruler’. Considering the fact that during evolution only 5 fireworks are used, it is understandable why the algorithm converges to a local solution. In fact, because the initial positions are randomly chosen, for repeated executions of IR procedure on ‘Testpat’ image, in about 10% of cases the algorithm leads to the correct solution.

a. Histogram of ‘Lena’

b. Histogram of ‘Peppers’

c. Histogram of ‘Ruler’

d. Histogram of ‘Testpat’

Fig. 4. Histograms of the model images

516

S.-I. Bejinariu et al.

a. NMI variation – ‘Lena’

b. NMI variation – ‘Peppers’

c. NMI variation – ‘Ruler’

d. NMI variation – ‘Testpat’

Fig. 5. Normalized mutual information variation

The easiest solution to solve these failures is to change the algorithm’s parameters. The attempt to increase the number of iterations (epochs) to 500 or 1000 did not lead to the expected result, because it depends on the spatial distribution of the initial population. By increasing the number of fireworks in the initial population to 10 or 15, the percent of cases in which the algorithm leads to the correct solution has increased to about 25%. In our opinion, both these parameters must be increased, even if this solution is not feasible because the execution time increases also. Another solution of this problem is the usage of features based IR. Some results are shortly described later in the fifth section. 3.2

Image Registration of Noise Affected Images

The behavior of the FWA based IR procedure was studied also for noise affected images. The source images were warped using the same affine transform described in the previous section and then they were degraded using Gaussian or Salt & Pepper noise. The degraded source images were registered to the original model images. Image degradation is evaluated using the Signal-To-Noise ratio (Table 3). The objective evaluation of IR results is presented in Tables 4 and 5 and the processed images are presented in Fig. 6. The conclusion that can be drawn is that the registration process is not affected by the degradation of images, but depends on the similarity measure used as objective function in the optimization. In fact this demonstrates that the procedure can be applied for multimodal images.

Fireworks Algorithm Based Image Registration

517

Table 3. SNR value for degraded source images Model image SNR (dB) S&P noise Gaussian noise Peppers 2.46 21.19 Testpat 2.72 21.72

Table 4. Evaluation of IR of degraded images Image

Noise

NMI Expected Peppers Gaussian 1.2081 S&P 1.1160 Testpat Gaussian 1.2526 S&P 1.1407

Computed 1.2081 1.1166 1.2533 1.1410

NCC Expected 0.9662 0.9069 0.9513 0.8921

Computed 0.9661 0.9070 0.9517 0.8919

# evals

Best eval

Time (sec)

14289 14491 14516 14408

13812 14390 14497 13962

29.890 33.291 24.258 26.801

Table 5. Computed parameters of the inverse transform Image Noise Peppers Gaussian S&P Testpat Gaussian S&P

cx 29.8557 30.0255 30.5535 29.9779

cy 29.9494 30.0632 30.0995 30.0260

Angle −10.0009 −10.0169 −10.0220 −9.9906

Scale 1.2498 1.2504 1.2513 1.2497

a. Source: ‘Peppers’ + Gaussian noise

b. Source: c. Source: ‘Peppers’ + S&P noise ‘Testpat’ + Gaussian noise

d. Source: ‘Testpat’ + S&P noise

e. Registered: ‘Peppers’ + Gaussian noise

f. Registered: g. Registered: ‘Peppers’ + S&P noise ‘Testpat’ + Gaussian noise

h. Registered: ‘Testpat’ + S&P noise

Fig. 6. IR results obtained using degraded source images

518

S.-I. Bejinariu et al.

4 FWA Versus Other Optimization Algorithms The accuracy of the image registration results is compared to that obtained using other optimization algorithms: Particle Swarming (PSO), Cuckoo Search (CSA) and Genetic Algorithm (GA). The PSO implementation described in [13] was customized for IR with the following parameters: number of particles – 100, number of iterations – 100, inertia weight – 0.729, local weight – 1.4945 and global weight – 1.4945. The weights are those proposed in [16]. Our experiments demonstrated that even small variations of these values decrease the results accuracy. The CSA based IR procedure [12] was used with the following parameters: number of nests – 25, number of iterations – 1000 and discovery rate – 0.25. For GA based IR an implementation based on real encoding (each chromosome is characterized by real values representing the geometric transform parameters) was used [17]. Discrete, average and simplex crossover operators were applied with different probabilities. The GA was applied as in [17] using the following parameters: number of generations – 500, number of chromosomes – 1500, probability of discrete crossover – 0.05, probability of average crossover – 0.15, probability of simplex crossover – 0.2 and mutation rate – 0.2. The IR procedure was applied using all these optimization algorithms on the same images as in the previous section. The obtained results are described in Table 6. Both NMI and NCC values are presented but NMI was used as objective function to be maximized. The maximum values of NMI are embossed for each image case. Table 6. Evaluation of IR results using FWA, PSO, CSA and GA as optimization algorithm Image Expected NMI NCC Lena 1.3252 0.9765 Peppers 1.3463 0.9675 Ruler 1.1147 0.9699 Testpat 1.5014 0.9525

FWA Computed NMI NCC 1.3253 0.9766 1.3466 0.9675 1.0492 0.3511 1.5014 0.9525

PSO Computed NMI NCC 1.3256 0.9766 1.3472 0.9676 1.0199 0.3184 1.5015 0.9525

CSA Computed NMI NCC 1.3257 0.9766 1.3472 0.9676 1.0535 0.8995 1.5017 0.9526

GA Computed NMI NCC 1.3244 0.9765 1.3455 0.9675 1.0408 0.2925 1.4907 0.9523

The same results are graphically presented for each test image in Fig. 7. The values presented in Fig. 7 have been rounded to three decimal digits. For this reason, in some graphs the values are equal but the corresponding bars have different heights. A brief analysis of these results leads to the following conclusions: – First of all, it must be noticed that in case of ‘Ruler’ test image the registration fails also in case of PSO, CSA and GA optimization, – All four optimization algorithms produce enough good results for three test images: ‘Lena’, ‘Peppers’ and ‘Testpat’. – CSA offers the best results for all four images. In fact, also for ‘Ruler’ test image, FWA, PSO and GA produce a result similar to that presented in Fig. 3c, while CSA

Fireworks Algorithm Based Image Registration

a. ‘Lena’ image

b. ‘Peppers’ image

c. ‘Ruler’ image

d. ‘Testpat’ image

519

Fig. 7. Evaluation of pixel based IR results obtained using FWA, PSO, CSA and GA for each test image

tends to approximate the inverse geometric transform. The ‘scale’ and ‘angle’ parameters are enough well approximated, only the rotation center coordinates are pretty much shifted from the expected values. It is possible that an increased number of iterations and/or an increased number of host nests allow CSA to obtain a better solution. – The results obtained by using GA in optimization are less accurate than the other three algorithms results. – The accuracy of results obtained by using PSO and FWA is good even if lower than CSA results accuracy. The PSO, CSA and GA based IR procedures were applied also for the images altered by Gaussian and Salt & Pepper noise with quite similar results. A complete analysis of the registration results have to consider also the processing time which is determined by the number of objective function evaluations. As it was already noticed in the previous section, in case of pixel based IR applications, more than 90% of the processing time is spent in these evaluations, so, reducing them may be important at least for real time applications. In Fig. 8, the average number of objective function evaluations is presented. On the other hand, the number of evaluations required to obtain the best solution is important given the fact that this class of algorithms usually obtain the best solution during the evolution but not necessary in the last iteration. As is depicted in Fig. 8, FWA and PSO use a lower number of cost function evaluations, while GA uses the greatest number while its results are less accurate. Concerning the number of cost function evaluations required to obtain the best solution,

520

S.-I. Bejinariu et al.

Fig. 8. Average number of objective function evaluations

it must be noticed that in case of FWA, PSO and CSA it is close to the total number of evaluations. This leads to the conclusion that more precise solutions can be obtained by increasing the number of iterations and/or individuals used in the evolutionary process. A different situation is encountered in case of GA. Its best solutions are obtained after about 8–9000 cost function evaluations, i.e. about one eight of the total number which is lower than in case of the other algorithms. This means that GA optimization represents a good choice if less accurate solutions are acceptable. As a conclusion, CSA can be used to obtain more precise solutions; GA can be used to faster obtain an acceptable solution; FWA and PSO is the proper choice if good solutions have to be also quickly obtained.

5 Pixel Based Versus Features Based IR We still have to find a solution to register images that contain graphic drawings, like the ‘Ruler’ test image. A good solution is the usage of features based IR. In this case, distinctive and stable features (points, lines, contours, regions) have to be detected in both model and source images. A features based IR procedure using bio-inspired computing is presented in [17]. As features the key-points determined using the Scale Invariant Feature Transform (SIFT) [18] are used. To evaluate images similarity, the correspondences between features present in both images have to be identified. The Euclidean distances between SIFT descriptors are computed for all key-points pairs detected in the model and source images. For each key-point descriptor in the model image, the distances to the key-points descriptors in the source image are sorted in ascendant order. If the smallest distance is less than a specific percent (usually 30%) of the second distance then a match is established between the key-point in the model image and the key-point in the source image [17]. The registration procedures are similar excepting the objective function used in optimization. In case of features based IR the Euclidean distance between the positions of the common features in the two images is used for images similarity evaluation. It must be noticed also that in this case the geometric transform corresponding to each

Fireworks Algorithm Based Image Registration

521

possible solution have to be applied only for the key-points coordinates while in the pixel based IR, the entire source image have to be transformed. This explains why the features based IR is much faster than pixel based IR. In fact, increasing the number of iterations and/or individuals does not have a great impact on the processing time. However, features based IR has one great disadvantage. For multimodal or some noise affected images the correspondences between stable features in the two images can’t be established. Trying to register images that contain graphic drawings (‘Ruler’ image), the four optimization algorithms were applied for features based IR using the SIFT key-points. A number of 2607 stable features were determined in the model image and 2525 stable features were determined in the source image. From these only for 60 pairs of key-points the correspondence was established and they were used in registration. The image registration succeeded for all algorithms. As depicted in Fig. 9, the best registration accuracy is obtained when FWA was used as optimization algorithm but also the others produce enough good results. Concerning the number of objective function evaluations it is similar to that of pixel based IR, because the parameters chosen for the optimization algorithms were the same. Anyway this is less important because the IR is very fast. As comparison, the duration of IR procedure applied for the ‘Lena’ image and FWA usage in optimization, on a system equipped with an Intel Core i5 3.1 GHz processor and 4 GB RAM, is shown in Table 7.

Fig. 9. Evaluation of features based IR results using FWA, PSO, CSA and GA for ‘Ruler’ test image Table 7. IR duration (seconds) for ‘Lena’ image and FWA optimization Pixel based IR Features based IR Parallel 14.72 0.04 Sequential 20.13 0.05

The features based IR was applied also for the images degraded using Gaussian and Salt & Pepper noise. In the first case the features based IR succeeded as well even if fewer correspondences were found in the two images. In the second case, the

522

S.-I. Bejinariu et al.

registration failed because no correspondences were established between the features detected in the two images.

6 Conclusions In this paper, the behavior of the Fireworks optimization algorithm for image registration problems is analyzed. Both pixel based and features based image registration were addressed for different types of images: pictures, graphics and a combination of these. In most cases the optimization algorithm is chosen and its parameters are tuned depending on the problem to be solved. FWA is compared to other nature inspired evolutionary algorithms: Particle Swarming, Cuckoo Search and Genetic algorithms. Concerning the pixel based IR, for which the Normalized Mutual Information was used as objective function, FWA may be a choice if enough good registration results have to be obtained in a reasonable execution time because the accuracy of its results is almost as good as for PSO and CSA algorithms and better than those of GA usage. Regarding the number of objective function evaluations that is proportional to the execution time, in the studied cases FWA is almost as fast as PSO is and much faster than CSA and GA. Similar results were obtained for images degraded by noise. The pixel based registration may fail if the images to be registered contain graphics. For this reason a features based IR approach is also analyzed. It is based on the SIFT features and FWA offers more accurate results than PSO, CSA and GA optimization algorithms. The features based IR has the disadvantage that it might fail for multi-modal or noise affected images, because the correspondences between the stable features detected in the two images can’t be found. The results presented in this paper were obtained using FWA, PSO, CSA and GA implementations in C++ [9–14]. The parallel implementation uses the computing power of multi-core processors and was developed using the parallel computing support of the Microsoft Visual Studio 2015 framework. The OpenCV open-source library was used for images manipulation. The graphs presented in Fig. 5 were drawn using a Matlab application. This research will be continued by the analysis of other nature-inspired optimization algorithms and their usage in image processing applications.

References 1. Zitova, B., Flusser, J.: Image registration methods: a survey. Image Vis. Comput. 21, 977– 1000 (2003). Elsevier 2. Yang, X.-S.: Nature-Inspired Optimization Algorithms. Elsevier Inc., Amsterdam (2014) 3. Tan, Y., Zhu, Y.: Fireworks algorithm for optimization. In: Tan, Y., Shi, Y., Tan, K.C. (eds.) ICSI 2010, Part I. LNCS, vol. 6145, pp. 355–364 (2010) 4. Tan, Y.: Fireworks Algorithm A Novel Swarm Intelligence Optimization Method. Springer, Heidelberg (2015)

Fireworks Algorithm Based Image Registration

523

5. Zheng, S., Janecek, A., Tan, Y.: Enhanced fireworks algorithm. In: Proceedings of 2013 IEEE Congress on Evolutionary Computation, Cancún, México, pp. 2069–2077 (2013) 6. Li, J., Zheng, S., Tan, Y.: Adaptive fireworks algorithm. In: Proceedings of 2014 IEEE Congress on Evolutionary Computation (CEC), pp. 3214–3221 (2014) 7. Zheng, S., Li, J., Janecek, A., Tan, Y.: A cooperative framework for fireworks algorithm. In: IEEE/ACM Transactions on Computational Biology and Bioinformatics, pp. 1–13 (2015) 8. Liu, L., Zheng, S., Tan, Y.: S-metric based multi-objective fireworks algorithm. In: 2015 IEEE Congress on Evolutionary Computation (CEC), pp. 1257–1264 (2015) 9. Costin, H., Bejinariu, S.I.: Medical image registration by means of a bio-inspired optimization strategy. Comput. Sci. J. Moldova 20, 2(59), 178–202 (2012) 10. Bejinariu, S.-I.: Image registration using bacterial foraging optimization algorithm on multi-core processors. In: 4th International Symposium on Electrical and Electronics Engineering (ISEEE), Galaţi, România (2013) 11. Bejinariu, S.-I., Costin, H., Rotaru, F., Luca, R., Niţă, C.: Social behavior in bacterial foraging optimization algorithm for image registration. In: Proceedings of the 18th International Conference on System Theory, Control and Computing, Sinaia, Romania, pp. 330–334 (2014) 12. Bejinariu, S.-I., Costin, H., Rotaru, F., Luca, R., Nita, C.D.: Image processing by means of some bio-inspired optimization algorithms. In: Proceedings of the IEEE 5th International Conference on E-Health and Bioengineering – “EHB 2015”, Iasi, Romania, pp. 1–4 (2015) 13. Bejinariu, S.-I., Costin, H., Rotaru, F., Luca, R., Nita, C.D.: Automatic multi-threshold image segmentation using metaheuristic algorithms. In: 2015 International Symposium on Signals, Circuits and Systems (ISSCS), Iasi, Romania, pp. 1–4 (2015) 14. Bejinariu, S.-I., Costin, H., Rotaru, F., Luca, R., Nita, C.D.: Fireworks algorithm based single and multi-objective optimization. Paper Submitted to Buletinul Institutului Politehnic din Iasi, Sectia Automatica si Calculatoare (2016) 15. University of Southern California, USC-SIPI Image Database. http://sipi.usc.edu/database/ database.php?volume=misc. Accessed 15 Mar 2016 16. Pedersen, M.E.H.: Good parameters for particle swarm optimization. Hvass Laboratories, Technical report no. HL100 (2010) 17. Bejinariu, S.-I., Costin, H., Rotaru, F., Luca, R., Niţă, C., Lazăr, C.: Parallel processing and bio-inspired computing for biomedical image registration. Comput. Sci. J. Moldova 22, 2(65), 253–277 (2014). Invited Article 18. Lowe, D.: Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis. 60(2), 91–110 (2004)

Discrete Wavelet Transforms for PET Image Reduction/Expansion (wavREPro) Hedi Amri1(B) , Malek Gargouri1 , Med Karim Abdmouleh1 , Ali Khalfallah1 , Bertrand David2 , and Med Salim Bouhlel1 1

Research Unit: Sciences and Technologies of Image and Telecommunications, Higher Institute of Biotechnology, University of Sfax, Sfax, Tunisia {hedi.amri,malek.gargouri}@setit.rnu.tn, [email protected], [email protected], [email protected] 2 LIRIS Laboratory, UMR 5205 CNRS Central School of Lyon, University of Lyon, 36, av Guy de Collongue, 69134, Lyon-Ecully Cedex, France [email protected]

Abstract. The large volume of medical images remains a major problem for their archiving and transmission. In this context, we propose a novel protocol wavREPro that aims to minimize the image size before its storage and then to enlarge it at reception. This process has been achieved by exploiting the Discrete Wavelet Transforms (DWT) namely Haar, Le Gall (5/3) and Daubechies (9/7) wavelets. These tools represent the image in the multi-resolution domain that follows the human psycho-visual system. Therefore, the reduced image is none other than the approximation of the image. Its magnification is carried out by either cancelling the details (wavREProZ) or estimating them (wavREProED) using the DWT−1 on the reduced image. Our experiments have been conducted on a PET (Positron Emission Tomography) medical image database and the results have been presented for the three well-known color spaces RGB, HSV and YCbCr. The reported results have promoted the wavREProZ application with the Haar wavelets on RGB images since it achieved maximum fidelity between the original and reduced then enlarged images. The good performance of this approach encourages its adoption to display images on screens having different sizes. Keywords: Telemedicine · Archiving · Transmission · Medical images · Discrete Wavelet Transforms · Image reduction · Image expansion

1

Introduction

The reduction and expansion of digital images are current, useful and even necessary operations in many applications related to the imaging domain especially in the medical field [7,14]. In fact, the big size of medical images [4,24] make their transmission and archiving quite difficult due to the limitation of both the hard drive capacity and the transfer rate [10]. The only possible solution to overcome this problem is to reduce the size of these files by compression techniques c Springer International Publishing AG 2018  V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8 45

Discrete Wavelet Transforms for PET Image Reduction/Expansion

525

such as JPEG [11], JPEG 2000 [34], Fractal coding methods [6], Region of Interest Coding Techniques [12,17], Lossless dynamic and adaptive compression [9], low-complexity compression [29] and genetic algorithms [33] and REPro (Reduction/Expansion Protocol) [2]. In this work, we are interested in REPro [11] that compresses images by reducing their definition (pixel number). In this context, various methods have been developed such as the decimation by square-square mesh [15], the decimation from a square mesh to a staggered mesh [28] and the decimation from a staggered mesh to a square mesh. On the other hand, the enlargement process is required to access to the image details. It can also be used to increase the resolution of the image in the same spatial support to improve the visual comfort of the viewer. In the literature, different techniques have been utilized for enlarging color images [27] namely the zero-padding [26], the polynomial nearest neighbor [32], the linear, quadratic and cubic interpolations and the cardinal spline transformation [13]. In our research, a novel approach has been developed to reduce and enlarge PET (Positron Emission Tomography) images. We have exploited the image multiresolution representation using Discrete Wavelet Transforms (DWTs) [8, 16,19,25,30] in the reduction and expansion phases. This paper is organized as follows. Section 2 is reserved to the presentation of our new approach wavREPro and its two versions. In this part, we also present an overview about the different used tools which are the DWTs and the image quality metrics. Section 3 presents our experimental results and their assessments. In the last section (Sect. 4), we summarize the main contributions of our work.

2 2.1

Proposed Approach Medical Context

Telemedicine is an effective solution to remedy the lack of medical specialists and equipment. Indeed, thanks to this technology, it is possible to exchange information having different sizes and natures between a tele-staff. This fact requires the adaptation of the image size to the screen size mainly when displaying medical images [20]. However, these images are usually large which can slow the communication. Furthermore, to display these images, they must also be adapted to the resolution of the screens [31]. In this context, REPro permits the reduction of the image definition that speeds its transmission. On the other hand, this protocol can enlarge the transferred image to ensure more visual comfort or it may further reduce the image definition to be adaptable to the low resolutions of some display devices such as tablets and smart phones (Fig. 1). 2.2

Methodology

To reduce or to enlarge the image, we have chosen to transform the image to the multiresolution domain to benefit from its simultaneous spatial and frequency locations. This process is carried out by the use of the Discrete Wavelet Transforms (Fig. 2).

526

H. Amri et al.

Fig. 1. Image transmission process using REPro

Fig. 2. Protocol wavREPro

Where “ori.Im” is the original image, DWT is the Discrete Wavelet Transforms, LUT is the Look-Up Table, RI is the reduced image, RILUT is the approximation space related to the applied DWT, REI is the reduced then enlarged image. 2.2.1 Used Wavelet Transforms Wavelets are interesting tools as they provide a simultaneous frequency and spatial locations of the image. This feature following the human psychovisual model allows wavelets to be exploited in various image processing applications such as compression (JPEG 2000), filtering [3], segmentation [5] and watermarking [19]. To apply a DWT on an image, we simply utilized it on lines then on columns. The original image was first decomposed into an approximation Ap0 and 3 detail sub images called HD0 , VD0 and DD0 which denoted Horizontal Details, Vertical Details and Diagonal details, respectively [23]. The recursive application of wavelets yielded the representation of the image in the multiresolution domain at level n as shown in Fig. 3. The level n denoted the number of iterations of the application of wavelets.

Discrete Wavelet Transforms for PET Image Reduction/Expansion

527

Fig. 3. Representation of an image decomposition: one level and two level

In this work, three well-known transforms were exploited. In fact, the Haar wavelet was used for its simple implementation [22]. Moreover, we utilized Le Gall 5/3 and Daubechies 9/7 wavelets due to their frequent usages in the JPEG 2000 compression standard. The equations of Haar, 5/3 et 9/7 DWTs are depicted in Eqs. 1, 2 and 3, respectively.  s[n] = 12 (s0 [n] + d0 [n]) (1) d[n] = 12 (s0 [n] − d0 [n])    d[n] = d0 [n] - 12 (s0 [n + 1] + s0 [n]) (2)   s[n] = s0 [n] + 14 (d[n] + d[n - 1]) + 12  1  ((s0 [n + 2] + s0 [n − 1]) − 9 (s0 [n + 1] + s0 [n])) + 12 d[n] = d0 [n] + 16   s[n] = s0 [n] + 14 (d[n] + d[n - 1]) + 12 (3) where s0 and d0 are the initial approximation and details and s and d are the new approximation and details. Figure 4 illustrated the image decompositions in one level and two level after the application of the 5/3 wavelet on an original color image (Fig. 4a).

Fig. 4. Image decomposition using the 5/3 wavelet: (a) original image, (b) decomposition in level 1, (c) decomposition in level 2

528

H. Amri et al.

2.2.2 The WavREPro Protocol The wavREPro protocol (or “Reduction/Expansion Protocol based on wavelets”) exploited the DWTs not only to reduce the image but also to enlarge it. Two versions of this protocol were proposed (Fig. 5). On the one hand, we obtained wavREProZ when the wavREPro protocol used Zero details. On the other hand, wavREProED represented wavREPro using enlarged details. These two approaches used the same reduction process that involved DWTs. This aimed to get the approximation on which we applied a Look-Up Table (LUT) [1] transform before its transmission to respect the image coding standards. We explain below the expansion process of wavREProZ and wavREProED.

Fig. 5. Image reduction by wavREPro

2.2.2.1 The wavREProZ Protocol When the user received the reduced image (RI), it applied on it a color palette (Look-Up Table: LUT). This table was based on linear functions to make the values in accordance with the approximation space related to the applied DWT. The resulting matrix was called RILUT. We associated to this matrix 3 zero matrices having the same dimensions as RILUT. These matrices were considered as matrices of Horizontal Details (HD = 0), Vertical Details (VD = 0) and Diagonal Details (DD = 0). Finally, DWT−1 was applied on the set {RILUT, HD = 0, VD = 0, DD = 0} to get the reduced then enlarged image using wavREProZ and we named it REI-wavZ (Fig. 6).

Fig. 6. Expansion of the reduced image using wavREProZ

Discrete Wavelet Transforms for PET Image Reduction/Expansion

529

Fig. 7. Decomposition of the reduced image using DWT

2.2.2.2 The wavREProED Protocol This protocol tried to find the vertical (VD), horizontal (HD) and diagonal (DD) details lost in the reduction phase. This process was carried out in two phases. In fact, the DWT was applied on the reduced image as illustrated in Fig. 7. Then, the DWT−1 is carried out on each of the Horizontal Details HDR, Vertical Details VDR and Diagonal Details DDR of the reduced mage (RI). To compute one of these components, the other components should be cancelled as described in Fig. 8. Thus, to estimate a detail component, we have enlarged its counter-parts obtained from the application of the DWT on the reduced image while respecting the frequency location of these details.

Fig. 8. Computation of the different details using wavREProED: (a) computation of HD, (b) computation of VD, (c) computation of DD

Finally, to get the image enlarged by the wavREProED protocol, we have adapted the reduced image RI to the approximation space related to the used DWT to obtain the same image RILUT used by wavREProZ. The magnified image (REI-wavED) is obtained by applying the DWT−1 on the set {RILUT, HD, VD, DD} as illustrated in Fig. 9. 2.2.3 Image Quality Metrics The image quality metrics were considered in our experiments to assess the level of distortion introduced by the reduction-expansion process. In this work, we used the Peak Signal to Noise Ratio (PSNR) and the Structural Similarity Index Metric (SSIM) to get an objective evaluation of the performance of the proposed approaches.

530

H. Amri et al.

Fig. 9. Expansion of the reduced image using wavREProED

2.2.3.1 Structural SIMilarity The Structural SIMilarity (SSIM) index is a method for measuring the similarity between two images where we focus on the structure image fidelity. The SSIM [18] is expressed by Eq. 4. SSIM (x, y) =

(2μx μy + C1 )(2σxy + C2 ) (μ2x + μ2y + C1 )(σx2 + σy2 + C2 )

(4)

C1 = (K1 L)2 , C2 = (K2 L)2 where μx and μx denote the average values of the images X and Y respectively, σx and σy are the standard deviations (the square root of variance) of X and Y σxy is the covariance of X and Y. L is the dynamic range of the pixel values (255 for the images encoded on 8 bits) and C1 and C2 are small positive constants (K1 and K2  1). At each pixel (i, j), a local SSIM (i, j) index is defined by evaluating the average, the standard deviation and the covariance on a local neighbor around that pixel (i, j). The overall quality of the image is measured by the Means SSIM index (MSSIM) which is expressed by Eq. 5. M SSIM =

1  SSIM (i, j) M i j

(5)

where M is the total number of local SSIM indices. 2.2.3.2 Peak Signal Noise Ratio The PSNR quantifies the image distortion in decibel (dB) [21]. It is a function of the Mean Square Error (MSE). Equations 6 and 7 describe the MSE and PSNR metrics, respectively. n  m    ∗ 2 Iij − Iij M SE =

i=1 j=1

(6)

nm 2

P SN R (I, I ∗ ) = 10 log 10(X max /M SE(I, I ∗ )) where I and I* are images having the same size n × m.

(7)

Discrete Wavelet Transforms for PET Image Reduction/Expansion

531

When a PSNR is upper than or equal to 30 dB, it reflects an acceptable image degradation. From 60 dB, it represents an imperceptible image distortion. 2.2.3.3 Selection Rate The quantification of the distortion introduced by two methods on an image database was generally carried out according results provided by P SN Raverage and SSIMaverage . However, if the P SN Raverage of the first approach is higher than that of the second method, this fact does not prove that the first approach introduces less distortion on each image of the database. To get a more objective assessment, we introduced the concept of selection rate. Hence, the selection rate of the method 1 was simply the percentage of the images of the database where this method introduced less distortion compared to the second method. Considering this definition, we obtained selection rates based on PSNR and SSIM to evaluate the performances of the approaches on our image database.

3

Experiments and Discussion

We proposed to evaluate the performance of our two approaches wavREProZ and wavREproED and to analyze the impact of each color space (RGB, HSV, YCbCr) and each adopted wavelet on the reduction-expansion process. Our experiments were conducted on 30 PET images. Initially, we were interested in assessing the performance of each wavREPro version when the images were represented in the RGB color space. The reported results in Fig. 10 and Table 1 obviously showed that the adoption of the Haar wavelet in wavREPro yielded two different performances n terms of PSNR and SSIM. Indeed, the use of this wavelet in the wavREProZ process, hence its name wavHaarZ, ensured the best degree of similarity between the original and reduced then enlarged images. However, the same wavelet caused the most image quality degradation when it was used in the wavREProED method, hence its name wavHaarED. In addition, we noted that PSNRs obtained after the application of 5/3 and 9/7 wavelets on PET images were very close (Fig. 10a). However, the results shown in Fig. 10b promoted the utilization of the 9/7 wavelet followed by the 5/3 wavelet in terms of SSIM. Thus, the Haar wavelet improved the performance of wavREProZ while it was preferable to use wavREProED with the 9/7 or 5/3 wavelets (Table 1). Where wav5/3Z and wav5/3ED denoted the 5/3 wavelets applied in wavREProZ and wavREProED, respectively. Similarly, wav9/7Z and wav9/7ED represented the 9/7 wavelets applied in wavREProZ and wavREProED, respectively. The results provided by Table 2 confirmed the values of PSNR and SSIM. Indeed, wavHaarZ was selected as the best technique of reduction-expansion for all the PET images of the database when the color space was RGB.

532

H. Amri et al.

Fig. 10. Effects of the wavREPro application on PET images represented in RGB color space: (a) Quality assessment of the RGB images in terms of PSNR, (b) Quality assessment of the RGB images in terms of SSIM

In order to evaluate the performances of wavREProZ and wavREProED and the impact of the choice of wavelets on the HSV (Hue Saturation Value) image quality, we converted the test images to this color space. Then, these images were reduced then enlarged. Finally, they were converted back to the RGB space. We exploited PSNR, SSIM and selection rates to illustrate the effect of the reduction operation followed by the image magnification process. The results presented in Fig. 11 and Table 3 proved again that wavReproZ based on Haar wavelet transforms performed better in terms of PSNR and SSIM. Similarly to the results attained in the RGB color space, the Haar wavelet had the greatest performance when we adopted it in wavREProZ. But, contrary to the results provided in the RGB color space, wavREProED based on wavHaarED seemed more efficient than wavREProED based on the 9/7 wavelets. Indeed, wavHaarED performed better than wav9/7ED for 20% of the database images in terms of PSNR and reached even 26.6% of this set in terms of SSIM (Table 4).

Discrete Wavelet Transforms for PET Image Reduction/Expansion

533

Table 1. Quality evaluation of the RGB images after the application of wavREPro. wavHaarZ wavHaarED wav5/3Z wav5/3ED wav9/7Z wav9/7ED P SN Raverage 50.47 SSIMaverage

35.42

0.989

38.21

0.877

0.904

39.54 0.922

41.51 0.959

42.63 0.97

Table 2. Selection rates of wavelets applied on RGB images. wavHaarZ wavHaarED wav5/3Z wav5/3ED wav9/7Z wav9/7ED PSNR 100%

0.0%

0.0%

0.0%

0.0%

0.0%

SSIM

0.0%

0.0%

0.0%

0.0%

0.0%

100%

Fig. 11. Effects of the wavREPro application on PET images represented in HSV color space: (a) Quality assessment of the HSV images in terms of PSNR, (b) Quality assessment of the HSV images in terms of SSIM

534

H. Amri et al.

Table 3. Quality evaluation of the HSV images after the application of wavREPro. wavHaarZ wavHaarED wav5/3Z wav5/3ED wav9/7Z wav9/7ED P SN Raverage 44.30 SSIMaverage

42.85

0.985

28.62

0.971

0.858

31.46 0.883

29.84

32.51

0.915

0.935

Table 4. Selection rates of wavelets applied on HSV images. wavHaarZ wavHaarED wav5/3Z wav5/3ED wav9/7Z wav9/7ED PSNR 80%

20%

0.0%

0.0%

0.0%

0.0%

SSIM

26.67%

0.0%

0.0%

0.0%

0.0%

73.33%

Fig. 12. Effects of the wavREPro application on PET images represented in YCbCr color space: (a) Quality assessment of the YCbCr images in terms of PSNR, (b) Quality assessment of the YCbCr images in terms of SSIM

Discrete Wavelet Transforms for PET Image Reduction/Expansion

535

We also carried out our experiments on our test images that we converted to the YCbCr color space. Figure 12 illustrated the degree of resemblance between the original and reduced then enlarged images. The performances of the different methods were summarized in Table 5. According to Fig. 11a, we clearly noticed that PSNRs are higher than 30 dB for all the applied methods. In other words, there was a good conservation of the content of all the images. Table 5. Quality evaluation of the YCbCr images after the application of wavREPro. wavHaarZ wavHaarED wav5/3Z wav5/3ED wav9/7Z wav9/7ED P SN Raverage 38.50

35.39

39.67

41.20

42.54

43.38

SSIMaverage

0.891

0.940

0.945

0.966

0.973

0.913

On the other hand, we noted that PSNR and SSIM reached great values when wav9/7ED was adopted (Table 5). Moreover, the 5/3 wavelet provided the second best performance followed by the Haar wavelet. Therefore, wav9/7ED was obviously selected as the best technique since it ensured the most image content preservation (Table 6). Table 6. Selection rates of wavelets applied on YCbCr images. wavHaarZ wavHaarED wav5/3Z wav5/3ED wav9/7Z wav9/7ED PSNR 0.0%

0.0%

0.0%

0.0%

0.0%

100%

SSIM

0.0%

0.0%

0.0%

0.0%

100%

0.0%

Finally, we noted that the Haar wavelet always gave the best results when it was included in wavREProZ whereas the 9/7 and 5/3 wavelets performed better when used in the wavREProED protocol. To highlight the impact of the space color choice on the wavREPro performance, we compared the best wavREPro approaches selected by each color space. This leaded us to compare the performances of wavHaarZ applied in RGB and HSV color spaces to those of wav9/7ED utilized in the YCbCr color space. Figure 13 and Tables 7 and 8 showed the performance assessments of methods in terms of PSNR and SSIM. We noted that RGB represented the best color space to adopt and that is why it was selected for all images (Fig. 13 and Tables 9 and 10).

536

H. Amri et al.

Fig. 13. Effects of the wavREPro application on PET images: (a) Quality assessment of the images in terms of PSNR, (b) Quality assessment of the images in terms of SSIM

Table 7. Performance assessment of selected wavelets applied on TEP images in terms of P SN Raverage . wavHaarZ (RGB) wavHaarZ (HSV) Ond9/7ED (YCbCr) 50.47

44.30

43.38

Table 8. Performance assessment of selected wavelets applied on TEP images in terms of SSIMaverage . wavHaarZ (RGB) wavHaarZ (HSV) wav9/7ED (YCbCr) 0.989

0.985

0.973

Discrete Wavelet Transforms for PET Image Reduction/Expansion

537

Table 9. Selection of the best wavelet according to PSNR. wavHaarZ (RGB) wavHaarZ (HSV) wav9/7ED (YCbCr) 100%

0.0%

0.0%

Table 10. Selection of the best wavelet according to SSIM. wavHaarZ (RGB) wavHaarZ (HSV) Wav9/7ED (YCbCr) 100%

4

0.0%

0.0%

Conclusion

In this paper, we presented a new approach of reduction and enlargement called wavREPro that exploited the image representation in the multiresolution domain. This field provided simultaneously the spatial and frequency locations of the image and followed the human psycho-visual model. The transition to this domain was achieved using the Discrete Wavelet Transforms. In this context, we used three types of DWTs namely Haar, Le Gall (5/3) and Daubechies (9/7) wavelets. These methods allowed the reduction of the image before its storage or transfer then its enlargement before its display on the screen. The minimized image was none other than the approximation that was the result of the DWT application on the image to transfer. To enlarge the image, we proposed two versions of wavREPro which added details to the obtained approximation and then applied the DWT−1 on the 4 components. The wavREPoZ version canceled details while the wavREProED version estimated the details of the image transformed to the multiresolution domain. The performances of these different approaches were assessed in different color spaces namely RGB, HSV and YCbCr. The obtained results promoted the use of the Haar wavelet, which canceled the details in the image magnification process. We also noted that it was appropriate to exploit the RGB space color for the application of wavREPro on PET images. In addition, it was obviously noted that the Haar wavelet was more effective when used in wavREProZ while wavREProED achieved good results when the 5/3 and 9/7 wavelets were utilized. This work has been exploited to reduce images before their storage or transfer. In addition, our approach has ensured their adaption to the different types of screens before their display. This can be carried out by resizing the image to fit appropriately the screen size.

References 1. Abdmouleh, M.K., Khalfallah, A., Bouhlel, M.S.: A chaotic cryptosystem for color image with dynamic look-up table. In: Proceedings of the 7th International Conference on Image and Signal Processing (ICISP 2016), pp. 91–100. Springer International Publishing, Trois-Rivieres (2016)

538

H. Amri et al.

2. Amri, H., Hanna, F., Lapayre, J.C., Khalfallah, A., Bouhlel, M.S.: REPRO: a new reduction/expansion protocol to increase the performance of image transmission in medical telediagnosis platforms. Biomed. Eng.: Appl. Basis Commun. 27(6), 1550054 (2015) 3. Aramendi, E., Irusta, U., Ayala, U., Naas, H., Kramer-Johansen, J., Eftestol, T.: Filtering mechanical chest compression artefacts from out-of-hospital cardiac arrest data. Resuscitation 98, 41–47 (2016) 4. Aribi, W., Khalfallah, A., Bouhlel, M.S., Elkadri, N.: Evaluation of image fusion techniques in nuclear medicine. In: 2012 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications (SETIT), pp. 875–880 (2012) 5. Bai, X., Jin, J.S., Feng, D.: Segmentation-based multilayer diagnosis lossless medical image compression. In: Proceedings of the Pan-Sydney Area Workshop on Visual Information Processing, VIP 2005, pp. 9–14 (2004) 6. Bhavani, S., Thanushkodi, K.G.: Comparison of fractal coding methods for medical image compression. IET Image Process. 7(7), 686–693 (2013) 7. Bonnans, V., Humbert, P., Pazart, L., Marzani, F., Lapayre, J.C., Lang, C.: Collaborative platform for skin cancer screening and associated optical fibered probe for diagnosis. In: Sin’Fran 2009, Singaporean-French IPAL Symposium, pp. 44–55 (2009) 8. Boudjelal, A., Messali, Z., Boubchir, L., Chetih, N.: Nonparametric Bayesian estimation structures in the wavelet domain of multiple noisy image copies. In: 2012 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications (SETIT), pp. 495–501 (2012) 9. Castiglione, A., Pizzolante, R., De Santis, A., Carpentieri, B., Castiglione, A., Palmieri, F.: Cloud-based adaptive compression and secure management services for 3D healthcare data. Future Gener. Comput. Syst. 43(44), 120–134 (2015) 10. Chaabouni, I., Fourati, W., Bouhlel, M.S.: Using ROI with ISOM compression to medical image. IJCVR 6(1/2), 65–76 (2016) 11. Ciznicki, M., Kurowski, K., Plaza, A.: Graphics processing unit implementation of JPEG2000 for hyperspectral image compression. J. Appl. Remote Sens. 6(1), 061507-1–061507-14 (2012) 12. Doukas, C., Maglogiannis, I.: Region of interest coding techniques for medical image compression. IEEE Eng. Med. Biol. Mag. 26(5), 29–35 (2007) 13. Fahmy, M.F., Fahmy, G., Fahmy, O.F.: B-spline wavelets for signal denoising and image compression. Signal Image Video Process. 5(2), 141–153 (2011) 14. Fuin, D., Garcia, E., Guyennet, H., Lapayre, J.C.: Collaborative interactions for medical e-Diagnosis. HPCN Int. J. High-Perform. Comput. Netw. 5(3), 189–197 (2008) 15. Gotchev, A.P., Egiazarian, K.O., Marchokov, G., Saramki, T.: A near least squares method for image decimation. ICIP 2, 929–932 (2003) 16. Hassen, W., Mbainaibeye, J., Olivier, C., Amiri, H.: New video encoder based on wavelet coefficients and motion compensation. In: 2012 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications (SETIT), pp. 440–447 (2012) 17. Hernndez-Cabronero, M., Blanes, I., Pinho, A.J., Marcellin, M.W., Serra-Sagrist, J.: Progressive lossy-to-lossless compression of dna microarray images. IEEE Signal Process. Lett. 23(5), 698–702 (2016) 18. Joshi, Y.G., Loo, J., Shah, P., Rahman, S., Chang, Y.C.: A novel low complexity local hybrid pseudo-SSIM-SATD distortion metric towards perceptual rate control. In: 2013 IEEE International Symposium on Broadband Multimedia Systems and Broadcasting (BMSB), pp. 1–6 (2013)

Discrete Wavelet Transforms for PET Image Reduction/Expansion

539

19. Kammoun, F., Khalfallah, A., Bouhlel, M.S.: New scheme of digital watermarking using an adaptive embedding strength applied on multiresolution filed by 9/7 wavelet. Int. J. Imaging Syst. Technol. 16(6), 249–257 (2006) 20. Kassab, R., Lapayre, J.C., Aupet, J.B., Marzani, F., Pieralli, C.: Scars collaborative telediagnosis platform using adaptive image flow. Integr. Comput.-Aided Eng. ICAE 1(20), 3–14 (2013) 21. Amri, H., Khalfallah, A., Lapayre, J.C., Bouhlel, M.S.: Watermarking for improving the reduction-expansion process of medical images (WREPro). Int. J. Imaging Robot.TM 16(3), 124–139 (2016) 22. Khalfallah, A., Bouhlel, M.S.: The impact of the concept of the family relative signatures on the non-blind watermarking in the multiresolution domain using 9/7 and 5/3 wavelets. Int. J. Inf. Commun. Technol. 4(3–4), 111–118 (2011) 23. Khalfallah, A., Brahim, K., Olivier, C.: A new approach to encrypted semi-blend watermarking in multi resolution field by 9/7 wavelet. In: IEEE International Conference of E-Medical Systems E-Medisys 2007. Frs, Morocco (2007) 24. Lazrag, H., Naceur, M.S.: Despeckling of intravascular ultrasound images using curvelet transform. In: 2012 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications (SETIT), pp. 365–369 (2012) 25. Lazrag, H., Naceur, M.S.: Wavelet filters analysis for speckle reduction in intravascular ultrasound images. In: 2012 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications (SETIT), pp. 375–379 (2012) 26. Liu, S., Wang, Q., Liu, G.: A versatile method of discrete convolution and FFT (DC-FFT) for contact analyses. Wear 243(1/2), 101–111 (2000) 27. Nini, B.: Projection based permutation of color images. In: 2012 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications (SETIT), pp. 460–467 (2012) 28. Petraglia, A., Pereira, J., Barugui, F.: Low-sensitivity IIR switched-capacitor decimation filters using delay lines with staggered T/H stages. IEEE Proc. - Circuits Devices Syst. 153(5), 193–198 (2006) 29. Pizzolante, R., Carpentieri, B., Castiglione, A.: A secure low complexity approach for compression and transmission of 3-D medical images. In: Proceedings of the 2013 Eighth International Conference on Broadband and Wireless Computing, Communication and Applications, BWCCA 2013, pp. 387–392. IEEE Computer Society (2013) 30. Saidani, T., Atri, M., Said, Y., Tourki, R.: Real time FPGA acceleration for discrete wavelet transform of the 5/3 filter for JPEG 2000. In: 2012 6th International Conference on Sciences of Electronics, Technologies of Information and Telecommunications (SETIT), pp. 393–399 (2012) 31. Sanchez Santana, M.A., Aupet, J.B., Betbeder, M.L., Lapayre, J.C., Camarena, A.: A tool for telediagnosis of cardiovascular diseases in a collaborative and adaptive approach. J. Univers. Comput. Sci. JUCS 19(9), 1275–1294 (2013) 32. Schafer, R.W., Rabiner, L.R.: A digital signal processing approach to interpolation. Proc. IEEE 61(6), 692–702 (1973) 33. Shih, F.Y., Wu, Y.T.: Robust watermarking and compression for medical images based on genetic algorithms. Inf. Sci. 175(3), 200–216 (2005) 34. Wu, X., Li, Y., Liu, K., Wang, K., Wang, L.: Massive parallel implementation of JPEG2000 decoding algorithm with multi-GPUS. In: Proceedings of the SPIE, vol. 9124, pp. 91240S–91240S-6 (2014)

Tips on Texture Evaluation with Dual Tree Complex Wavelet Transform Anca Ignat1(&) and Mihaela Luca2 1

Faculty of Computer Science, University “Alexandru Ioan Cuza” of Iași, Iași, Romania [email protected] 2 Institute of Computer Science, Romanian Academy, Iaşi, Romania [email protected]

Abstract. Dual Tree Complex Wavelet Transform (DTCWT) is a practical tool offering the possibility of extracting characteristics from images, a transformation which has good directional selectivity, is approximately shift invariant and it is computationally efficient. DTCWT may be employed in different stages of image processing, including multiscale texture featuring. We are measuring texture similarities on characteristics vectors by applying cosine, Pearson correlation coefficient, entropy-based and histogram related distances and we tested these measures on slightly rotated samples from Brodatz texture album. Comparisons and observations are made on the methods that we employed. Keywords: Dual Tree Complex Wavelet Transform Similarity measures
Texture




Feature extraction




1 Introduction Understanding images is a complex process which implies training during long periods of time in order to use high amounts of accumulated knowledge. To recognize images similar to those that we are acquainted to, we are scaling them and situating them in environmental contexts, in order to place the objects from the image in an appropriate scenario. As human beings we are merely evaluating colors, textures, lines, shapes, shades and lights, materials and spatial positioning, while a computer needs a large variety of mathematical models attempting to represent the cognitive processes in order to obtain comparable results. Intelligent agent architectures aim to adapt to each special task apparently simple for human eye-brain chain which in fact is a very complex paradigm. Image features are extracted and decomposed in hierarchical stages, color and texture being characterized by complex mathematical methods. Texture characterization is important in many domains, among them being the medical imaging evaluation and interpretation, industrial applications, remote sensing, etc. [1–4]. With the actual amazing increase of the internet image databases it is essential in automatic object retrieval to have a good texture identification procedure. The texture featuring and analysis with a relatively new mathematical tool, the Dual Tree Complex Wavelet Transform DTCWT [5, 6], is discussed in our paper. Note that the DTCWT is abbreviated with an italic C to differ from the Continuous Wavelet Transform. © Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8_46

Tips on Texture Evaluation with Dual Tree Complex Wavelet Transform

541

We used the classic Brodatz album with texture images [7] and some images from our experimental database. Texture features were computed using the six directional coefficients provided by DTCWT applied on all these texture images and their rotated versions, as we will further explain. In order to identify a certain texture in an area of an image, we apply several distances (cosine, Pearson correlation coefficient, entropy-based and histogram related) on the extracted features.

2 Disadvantages of Wavelet Transforms The main disadvantages of (real) wavelet transform [8, 9] are: oscillations around singularities (1), shift variance (2), aliasing (3) and the lack of directionality (4). These issues can be overcome by using Complex Wavelets. Yet, it is difficult to compute complex wavelets satisfying the perfect reconstruction property. In [5, 6], Kingsbury introduced the Dual Tree Complex Wavelet Transform (DTCWT) which is a new way of computing complex wavelets with perfect reconstruction with a 2d:1 redundancy for d dimensional signals. Important aspects about computing DTCWT, its properties and possible applications were developed in [9–11].

3 Discussion on Dual Tree Complex Wavelet Transform The decimated Discrete Wavelet Transform [12] has two major inconveniences: shift variance (small shifts in the input signal can generate big variations in the wavelet coefficients) and reduced directional selectivity (especially in diagonal directions). To overcome these limitations one can use complex wavelets. The difficulty in computing complex wavelets is given by the problem of designing perfectly reconstructed filter banks beyond the first level of decomposition. The problem was solved by Kingsbury [5, 6]. He developed the Dual Tree Complex Wavelet Transform, which provides approximate shift invariance, good directional selectivity, uses short linear phase filters with perfect reconstruction and has efficient computation costs (O(N) for 1D-signals). The toll to be paid for these properties is increased redundancy (2:1 for 1D-signals, 4:1 for 2D signals, 8:1 for 3D, and so on) but still it is much less than the undecimated wavelet transform. In order to obtain approximate shift invariance for the decimated DWT, one must double the sample rate at each level of decomposition; this is equivalent with having two parallel fully decimated trees and also the filters at level 1 must be delayed by one sample. After level 1, to reestablish the uniform interval sampling, one must work with odd length filter on one tree and even length filter on the other. In order to have a better symmetry between the trees odd and even filters must alternate when changing the decomposition level. The output obtained from the two parallel trees can be interpreted as the real and the imaginary parts of the complex wavelet coefficients. In order to compute CWT for 2D signals, each level of decomposition contains two stages: one consists in filtering the rows and the second stage is to filter the result along columns.

542

A. Ignat and M. Luca

When applied to digital images the DTCWT yields at each level of decomposition six bandpasses sub-images of complex coefficients. These sub-images provide strong directional selectivity at angles ±15º, ±45º, and ±75º. For feature extraction the energy, the mean, the standard deviation, the entropy of the six complex directional coefficients (or their amplitude) yielded by the DTCWT are usually employed: in [13, 14] they use energy and standard deviation, [15] works with the mean, the standard deviation of the amplitude of the sub-band images, and [16] employs the mean and the entropy. In [17] features were extracted using a threshold for counting the significant elements of the DTCWT coefficients. In [8, 18, 19], rotation invariant features were designed using DTCWT. In our paper, we study the influence of rotation and of changing the brightness of images (by applying a c transformation) over the DTCWT coefficients. As features we use the histograms of the DTCWT directional sub-images. For comparisons, we employed ten similarity measures for probability density functions presented in [20] and for classification we used the minimum-distance classifier and SVM (Support Vector Machine) [21].

4 Texture Features Selected with DTCWT Due to the exponentially growing amount of digital images, each new aspect has to be taken into account to facilitate the automatic image classification and clustering. In order to group the photos in classes depicting the same scene, it is possible to extract by means of segmentation some texture samples from each image. We might classify the image to a certain class by taking into account the information provided by the retrieved dominant textures. In order to identify textures, we extract feature vectors and we match the texture from the image with textures from a database. We tested some similarity measures in order to compare the textures of the test images with those from the database. The class in which we introduce the test image is the one that gives most matches. We study in this paper the effect of rotation on classification process using DTCWT for feature extraction and we proposed some similarity measures. As an example, Fig. 1 is presenting an image from our town which contains as main textures bark, grass and pavement (Fig. 2). In Fig. 3 we have the same scene but it was taken by rotating the photo camera with an angle about 15–20°. We manually extracted three samples of textures from both of the images. The selected textures are depicted in Figs. 2 and 4. The samples represent the same three textures from both images (bark, grass, and pavement) but they are not cropped from the same spot in order to catch different illuminations of the scene. Alike visual detection of natural environments, we note that analysing the directions of important lines, shades, colours, we may detect, on different scales, different properties that help us to identify a texture. Because of its good directional selectivity, we considered the DTCWT as an appropriate tool for texture characterization. We have considered a database containing 12 grayscale texture images from the Brodatz album (Fig. 5). We uniformly preprocessed these images (the original and the rotated ones) in order to obtain a size of 64  64.

Tips on Texture Evaluation with Dual Tree Complex Wavelet Transform

Fig. 1. Image of a sidewalk in Iasi, Romania

(a)

(b)

(c)

Fig. 2. (a), (b) Selected textures: bark, grass, pavement

Fig. 3. Rotated scene of the original image from Fig. 1

543

544

A. Ignat and M. Luca

Figure 6(a) and (b) are presenting the six directional DTCWT sub-band images Sj;1 for the texture samples from Fig. 4.

Fig. 4. Bark, grass, pavement samples from Fig. 3

Bark

Brick

Grass

Leather

Pigskin

Raffia

Sand

Straw

Water

Weave

Wood

Wool

Fig. 5. Brodatz textures database – 12 selected samples for our tests

For the DTCWT we used the following indexes for the six sub-images: j = 1 for +15º, j = 2 for +45º, j = 3 for +75º, j = 4 for −75º, j = 5 for −45º, and j = 6 for −15º. We tested different levels of wavelet decomposition (D = 1, 2, 3, 4) thus obtaining 2  6  D directional coefficients. The provided sub-band images are complex: Wj;d ¼ Vj;d þ iUj;d ; j ¼ 1; 2; . . .; 6:

ð1Þ


 The DTCWT provides the Vj;d ; Uj;d pairs for j = 1,…, 6, d = 1,…, D, along with the real lowpass image from the last decomposition
 level. For our computations we employed only the amplitude Sj,d of the Vj;d ; Uj;d coefficients, i.e.:   qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 þ U 2 ; j ¼ 1; . . .; 6; d ¼ 1; . . .; D Sj;d ¼ Wj;d  ¼ Vj;d j;d

ð2Þ

Tips on Texture Evaluation with Dual Tree Complex Wavelet Transform

545

(a)

(b)

Fig. 6. (a) Selected textures: bark, grass, pavement 2 and the DTCWT amplitudes in the three directions: 15˚, 45˚, 75˚. (b) Selected textures: bark, grass, pavement 2 and the DTCWT amplitudes in the three directions: −15˚, −45˚, −75˚, d = 1

546

A. Ignat and M. Luca


 We computed the Sj;d ; j ¼ 1; . . .; 6; d ¼ 1; . . .; D sub-band images for the normalized texture images (with intensity values in [0, 1]). We multiplied the Sj,d by 100, rounded to the nearest integer and computed the histogram, thus obtaining six 100–dimensional feature vectors for each wavelet decomposition level. In order to compare these histograms we tested some of the distances described in [20] (five from the v2 family, five from the Shannon’s entropy family, the cosine and the Pearson correlation coefficient). We selected five that gave the best results. Let p and q be two histograms of size m = 100 we want to compare. We used the following distances in order to compare the histograms: dcos ðp; qÞ ¼ 1  cosðp; qÞ

ð3Þ

m P

pi qi i¼1 s ffiffiffiffiffiffiffiffiffiffi ffi sffiffiffiffiffiffiffiffiffiffiffi cosðp; qÞ ¼ m m P P p2i q2i i¼1

dcorr ðp; qÞ ¼ r ðp; qÞ ¼

ð4Þ

i¼1

1  r ðp; qÞ 2

ð5Þ

covðp; qÞ rp rq

ð6Þ

where by cov(p,q) we denoted the covariance of the random variables p, q, and rp, rq are the standard deviations of p and q respectively. From the Shannon’s entropy family we used Topsøe, Jeffreys and Jensen difference distances: dTop ðp; qÞ ¼

m  X i¼1

2pi 2qi þ qi log pi log ðpi þ qi Þ ðpi þ qi Þ

dJ ðp; qÞ ¼

dJd ðp; qÞ ¼

ð7Þ

m X pi ðpi  qi Þ log qi i¼1

m  X pi log pi þ qi log qi i¼1



2



pi þ qi pi þ qi log 2 2

ð8Þ  ð9Þ

and from the v2 family we employed the squared v2: dSqChi ðp; qÞ ¼

m X ð pi  qi Þ 2 i¼1

pi þ qi

ð10Þ

Tips on Texture Evaluation with Dual Tree Complex Wavelet Transform

547

We employed these similarity measures in the following way: dhist ðX; Y Þ ¼

D X 6 X

d ðhistðSX;j;d Þ; histðSY;j;d ÞÞ

ð11Þ

d¼1 j¼1

where d* is one of the five distances described above. Figure 7 depicts the histograms of the DTCWT coefficients in the direction of 75° (histograms of S3) for the Sand, Water and Wood textures. As can be seen, these histograms can be used as a separation measure between images.

Fig. 7. Histogram of the DTCWT coefficients for the 75° direction for Sand, Water and Wood original textures

For testing the features and the classifiers we applied two types of transformations: a geometrical one (i.e. rotations with different rotation angles) and c intensity transforms. We used the following c transformation [22]: T ðr Þ ¼ r c :

ð12Þ

Using DTCWT, the above presented similarity measures and minimum distance classifiers or SVM we seek to retrieve for each transformed image the original texture that produced it.

5 Results and Discussions For the numerical tests we employed MATLAB. To compute the DTCWT we adapted [23] the software package from Professor Nick Kingsbury’s web page [24]. As mentioned above, we tested the extracted features and the classifiers on 12 textures selected from the Brodatz album ([7], Fig. 5). We consider as classifiers the minimum distance classifier (using the distances previously introduced, i.e. (3), (5), (7)–(10)) and SVM. All the texture images are of size 64  64.

548

A. Ignat and M. Luca

Four types of classes representing the textures were considered. The first type has classes containing only one texture, the original one. The rotated and the c transformed ones are used separately for tests. The last three types of classes contain more than one image. The original Brodatz album contains images of size 512  512. Using these images and resizing we computed rotated and c transformed textures of the same size (512  512). For each such texture we extracted 64 (8
8) non-overlapping sub-images of size 64  64. Each class that represents a texture is formed by putting together these 192 (3
64) images. We consider three situations: in the first two cases we treat independently the rotated and the c transformed images and in the third one the classes contain a mixture of all types of images (base, rotated, c transformed). In the rotation case the class is formed only with rotated sub-images and those extracted from the original texture. In the c transformed situation the class is formed only with original and c transformed images. We formed the training set by choosing randomly 50% of the class content (50% of each type of image: original, rotated, c transformed), the other ones were employed for classification purpose. We have rotated the texture images with the following angles: 5°, 10°, 15° and 30°. In Fig. 8 are depicted the Raffia and the Wood textures and their rotated versions.

Raffia.05

Raffia.10

Raffia.15

Raffia.30

Wood.05

Wood.10

Wood.15

Wood.30

Fig. 8. Brodatz textures database – 2 rotated samples for tests

We applied also several c transformations to the original Brodatz. In Fig. 9. we show for Grass, Wood and Sand different c-transformed variants of these textures. In this paper, we considered values for c in [0.5, 2]. In the case of classes with only one element we used the following five distances: cosine, correlation, Topsøe, Jensen difference, and squared v2 . For the rotation angle of 5°, 10° and 15° we get 100% identification rate for all five considered distances. For 30° rotated textures the cosine distance provides 3 errors, the correlation distance 4 errors, the other three distances have each 2 errors. The comparisons between the Raffia and Wood rotated textures and the original ones using the five similarity measures described applied to the histogram of the DT sub-band images are shown in Tables 1, 2, 3 and 4. The misidentifications appear for the 30° rotation angle.

Tips on Texture Evaluation with Dual Tree Complex Wavelet Transform

Grass

Grass γ=0.6

Grass γ=1.6

Grass γ=2

Wood

Wood γ=0.6 Wood γ=1.6

Wood γ=2

Sand

Sand γ=0.6

Sand γ=2

Sand γ=1.6

549

Fig. 9. c-transformed Brodatz textures – Grass, Wood and Sand Table 1. Cosine and correlation distances - rotated Raffia images Raffia.05 dcos dcorr Bark.00 0.295 0.170 Brick.00 0.333 0.181 Grass.00 0.328 0.183 Leather.00 0.155 0.081 Pigskin.00 0.160 0.085 Raffia.00 0.045 0.021 Sand.00 0.143 0.076 Straw.00 0.181 0.099 Water.00 0.281 0.147 Weave.00 0.228 0.119 Wood.00 0.342 0.178 Wool.00 0.136 0.072

Raffia.10 dcos dcorr 0.285 0.163 0.347 0.189 0.323 0.180 0.138 0.072 0.165 0.088 0.073 0.037 0.131 0.069 0.208 0.114 0.286 0.149 0.241 0.126 0.346 0.180 0.141 0.075

Raffia.15 dcos dcorr 0.280 0.161 0.342 0.186 0.324 0.182 0.125 0.065 0.175 0.094 0.080 0.040 0.122 0.064 0.193 0.106 0.284 0.149 0.204 0.106 0.337 0.175 0.145 0.077

Raffia.30 dcos dcorr 0.288 0.167 0.390 0.212 0.297 0.167 0.126 0.066 0.188 0.101 0.138 0.072 0.150 0.080 0.220 0.122 0.326 0.171 0.208 0.109 0.407 0.213 0.184 0.098

The cosine and the correlation distances make the same mistake (Leather) and the other three distances misidentifies Raffia with Sand. Both mistakes are understandable from the visual point of view taking into account that we use a directional instrument. The same thing happens with the Wood texture, i.e., the cosine and the correlation distances provide Weave as classification result and the other three classify Wood as Water. We repeated the same classification procedure as we used for the rotated textures for the c transformed ones. We get 100% identification rate for c 2 ½0:6; 1:7. In Table 5 are the classification results for Grass and Sand with c ¼ 0:6 and in Table 6 are the comparisons results for Grass and Wood with c ¼ 2.

550

A. Ignat and M. Luca Table 2. Topsøe, Jensen difference and squared v2 distances - rotated Raffia images Raffia.05 dJd dT Bark.00 12.7 6.4 Brick.00 29.0 14.5 Grass.00 18.6 9.3 Leather.00 12.0 6.0 Pigskin.00 9.8 4.9 Raffia.00 1.5 0.7 Sand.00 6.9 3.5 Straw.00 18.2 9.1 Water.00 30.8 15.4 Weave.00 16.5 8.3 Wood.00 29.2 14.6 Wool.00 9.5 4.7

dSq 21.0 48.8 30.9 20.4 17.8 2.7 12.8 30.3 50.8 27.3 46.5 16.9

Raffia.10 dT dJd 12.5 6.2 29.3 14.6 18.6 9.3 11.3 5.7 9.9 5.0 2.3 1.1 6.7 3.3 18.0 9.0 31.2 15.6 16.5 8.2 29.8 14.9 9.6 4.8

dSq 20.6 49.2 30.9 19.4 17.8 4.2 12.4 30.0 51.2 27.1 47.3 16.9

Raffia.15 dT dJd 12.3 6.2 29.5 14.8 18.3 9.2 11.0 5.5 9.8 4.9 3.0 1.5 6.5 3.2 17.8 8.9 31.4 15.7 15.6 7.8 30.4 15.2 9.8 4.9

dSq 20.5 49.7 30.7 18.9 17.6 5.6 12.0 29.8 51.5 25.8 48.4 17.4

Raffia.30 dT dJd 13.3 6.6 32.7 16.3 17.9 9.0 10.5 5.3 10.9 5.4 8.6 4.3 6.9 3.4 18.0 9.0 33.9 17.0 12.7 6.4 35.7 17.9 12.6 6.3

dSq 22.0 53.9 30.1 18.3 19.4 15.3 12.6 30.0 55.1 21.4 56.4 21.9

Table 3. Cosine and correlation distances - rotated Wood images Wood.05 dcos dcorr Bark.00 0.605 0.333 Brick.00 0.192 0.104 Grass.00 0.619 0.336 Leather.00 0.447 0.235 Pigskin.00 0.448 0.236 Raffia.00 0.350 0.182 Sand.00 0.421 0.221 Straw.00 0.409 0.220 Water.00 0.140 0.070 Weave.00 0.318 0.164 Wood.00 0.045 0.020 Wool.00 0.390 0.205

Wood.10 dcos dcorr 0.609 0.336 0.190 0.102 0.624 0.339 0.437 0.230 0.430 0.227 0.346 0.180 0.421 0.222 0.400 0.214 0.129 0.065 0.320 0.165 0.053 0.024 0.384 0.203

Wood.15 dcos dcorr 0.590 0.326 0.203 0.110 0.616 0.336 0.425 0.224 0.406 0.214 0.336 0.175 0.400 0.211 0.385 0.207 0.112 0.056 0.306 0.158 0.082 0.040 0.361 0.191

Wood.30 dcos dcorr 0.509 0.283 0.272 0.146 0.544 0.297 0.350 0.185 0.335 0.177 0.353 0.185 0.329 0.174 0.355 0.192 0.174 0.088 0.232 0.119 0.239 0.123 0.339 0.180

The five distances behave in the same manner as for the rotated textures: the cosine and the correlation misidentify the same texture, and Topsøe, Jensen difference and squared v2 make the same mistake. We also studied the influence of histogram equalization of the texture images before computing the DTCWT. When we use histogram equalization, we obtain better results for the c-transformed images i.e. the interval with zero errors in the classification process extends to c 2 ½0:05; 5:4. On the contrary when this procedure is performed on the rotated images the results are worse. In the case of histogram equalization for rotated textures classification the results are: the cosine and correlation distances have 1 error each for the 5°, 10°, 15° rotation angles and 4 respectively 5 misidentifications for 30°.

Tips on Texture Evaluation with Dual Tree Complex Wavelet Transform

551

Table 4. Topsøe, Jensen difference and squared v2 distances - rotated wood images Wood.05 dJd dT Bark.00 37.6 18.8 Brick.00 19.7 9.9 Grass.00 43.1 21.6 Leather.00 33.2 16.6 Pigskin.00 37.8 18.9 Raffia.00 30.0 15.0 Sand.00 35.4 17.7 Straw.00 28.1 14.0 Water.00 12.8 6.4 Weave.00 40.1 20.1 Wood.00 1.9 0.9 Wool.00 30.4 15.2

dSq 59.3 34.0 68.1 52.9 59.7 47.6 56.2 45.8 22.1 63.2 3.4 48.5

Wood.10 dT dJd 37.9 19.0 20.0 10.0 43.7 21.8 33.1 16.5 37.1 18.6 30.3 15.2 35.3 17.6 28.8 14.4 12.4 6.2 40.0 20.0 2.9 1.5 30.0 15.0

dSq 59.8 34.2 68.9 52.7 58.7 48.0 56.0 46.8 21.4 63.0 5.3 47.9

Wood.15 dT dJd 37.9 19.0 19.8 9.9 44.0 22.0 33.1 16.5 36.1 18.1 30.3 15.2 34.8 17.4 29.3 14.6 11.8 5.9 38.9 19.5 4.5 2.3 29.5 14.7

dSq 60.2 33.8 69.8 53.1 57.3 48.2 55.4 47.8 20.3 61.5 8.2 47.2

Wood.30 dT dJd 33.5 16.7 25.8 12.9 39.9 20.0 27.8 13.9 26.4 13.2 31.8 15.9 26.3 13.1 28.4 14.2 19.3 9.7 27.7 13.8 22.8 11.4 24.7 12.3

dSq 53.5 42.5 63.9 45.5 42.5 51.4 42.7 45.7 32.1 44.8 36.1 40.3

Table 5. Grass and Sand gamma transformed images comparisons (c ¼ 0:6) Grass dcos Bark 0.183 Brick 0.572 Grass 0.082 Leather 0.193 Pigskin 0.268 Raffia 0.287 Sand 0.251 Straw 0.240 Water 0.566 Weave 0.417 Wood 0.589 Wool 0.305

dcorr 0.112 0.318 0.047 0.109 0.148 0.159 0.141 0.136 0.310 0.231 0.318 0.170

dT 6.7 34.6 1.9 6.9 22.6 16.8 15.8 13.6 49.5 18.9 41.5 26.7

dJdiff 3.4 17.3 0.9 3.4 11.3 8.4 7.9 6.8 24.8 9.4 20.8 13.3

dSq 11.7 57.5 3.3 11.9 38.0 27.9 27.6 23.8 79.1 31.0 65.4 44.1

Sand dcos 0.338 0.289 0.399 0.202 0.138 0.135 0.118 0.223 0.207 0.180 0.364 0.114

dcorr 0.194 0.155 0.223 0.108 0.073 0.070 0.061 0.123 0.106 0.092 0.190 0.060

dT 18.5 23.4 28.6 17.9 3.6 11.1 4.0 21.6 22.3 13.1 35.5 3.9

dJdiff 9.3 11.7 14.3 9.0 1.8 5.6 2.0 10.8 11.1 6.5 17.8 1.9

dSq 31.5 39.1 48.8 31.2 6.3 20.7 7.4 36.4 36.5 22.1 57.2 6.8

The Topsøe, Jensen difference and squared v2 distances have 6 errors each for the 30° rotated textures and no misidentification for the smaller rotation angles. For the cases when the classes have more than one texture we used as classifiers SVM and minimum distance with cosine, Jeffreys, Topsøe, Jensen difference, and squared v2. When the rotations and c-transformations were treated separately, each class in the training set has 64 elements, 32 from the original texture and 32 from the transformed one. The test set contains a total of 768 images, 64 from each texture. The rotation angles are the same as previously (h 2 f5 ; 10 ; 15 ; 30 g) and the values for c are in f0:5; 0:6; 0:8; 0:81 ; 0:61 ; 2g. We observe from Table 7 that as h increases usually the classification results improve, one of the poorest results is for

552

A. Ignat and M. Luca Table 6. Grass and Wood gamma transformed images comparisons ( c ¼ 2) Grass dcos Bark 0.200 Brick 0.468 Grass 0.179 Leather 0.138 Pigskin 0.146 Raffia 0.171 Sand 0.134 Straw 0.205 Water 0.418 Weave 0.310 Wood 0.494 Wool 0.184

dcorr 0.119 0.256 0.102 0.075 0.078 0.092 0.072 0.115 0.224 0.168 0.262 0.099

dT 7.1 30.7 8.6 5.1 10.3 8.3 5.3 13.3 36.5 11.9 34.5 13.4

dJdiff 3.5 15.4 4.3 2.5 5.1 4.2 2.7 6.7 18.2 6.0 17.2 6.7

dSq 11.7 51.5 15.4 9.1 17.9 14.3 9.7 23.1 58.9 19.9 54.4 22.7

Wood dcos 0.526 0.187 0.530 0.349 0.386 0.286 0.384 0.291 0.162 0.344 0.112 0.320

dcorr 0.293 0.104 0.288 0.183 0.205 0.150 0.204 0.156 0.085 0.180 0.057 0.170

dT 32.1 17.4 36.3 26.8 41.0 30.6 37.3 18.7 23.1 43.2 6.9 33.7

dJdiff 16.0 8.7 18.1 13.4 20.5 15.3 18.7 9.3 11.5 21.6 3.5 16.9

dSq 51.0 31.1 57.2 42.9 64.5 48.1 59.7 30.4 38.6 67.7 12.3 53.7

Table 7. Classification results for rotated case

SVM Jeffreys Topsøe Jensen Squared v2 Cos

Rotation angles h ¼ 5 h ¼ 10 96.75% (25) 96.09% (30) 91.54% (65) 92.06% (61) 93.88% (47) 93.75% (48) 93.88% (47) 93.75% (48) 93.62% (49) 93.62% (49) 85.29% (113) 84.64% (118)

h ¼ 15 95.31% (36) 92.32% (59) 93.75% (48) 93.75% (48) 93.75% (48) 85.94% (108)

h ¼ 30 95.44% (35) 92.84% (55) 94.79% (40) 94.79% (40) 94.66% (41) 85.16% (114)

h ¼ 5 , the DTCWT provides better separation of textures if the rotation angle is greater than 10°. Best results are obtained with the SVM and the Topsøe, Jensen difference distances. When applying c transformations, we considered pairs ðg; 1=gÞ; g\1. As was expected as g approaches 1 the results are getting better. Comparing the results provided by each pair ðg; 1=gÞ; g\1 one remarks that we get better results for c values greater than 1 which provide better contrast than c < 1. Again, with the SVM one gets the best results and Topsøe, Jensen difference, and squared v2 distances come next, making the same number of mistakes (Table 8). We also analysed cases when all the textures were put together: first, the original and the rotated textures using all angles, second the original and all the c-transformed images, and the third was when all the images representing a texture were present in class content. In the first situation, each class has 160 elements and 1920 texture were tested. In the second case the classes have each 352 elements and 4224 textures were classified. In the last case, the classes contain 480 images each and 5760 texture were employed for classification. The results are in Table 9.

Tips on Texture Evaluation with Dual Tree Complex Wavelet Transform

553

Table 8. Classification results for c transformation case c Transformations c = 0.5 c = 0.6 SVM Jeffreys Topsøe Jensen Squared v2 cos

95.96% 95.70% 96.62% 96.62% 96.35% 87.89%

(31) (33) (26) (26) (28) (93)

97.66% 96.62% 97.79% 97.79% 97.79% 91.54%

(18) (26) (17) (17) (17) (65)

c = 0.8

c = 0.8−1

c = 0.6−1

c=2

99.74% (2) 96.62% (26) 100% (0) 100% (0) 100% (0) 95.44% (35)

99.74% (2) 100% (0) 99.87% (1) 99.87% (1) 99.87% (1) 98.57% (11)

97.92% 97.14% 96.75% 96.75% 97.01% 94.01%

96.22% 95.31% 95.83% 95.83% 95.83% 91.78%

(16) (22) (25) (25) (23) (46)

(29) (36) (32) (32) (32) (63)

Table 9. Classification results for rotations, c -transformed considered together SVM Jeffreys Topsøe Jensen Squared v2 cos

Rotations 97.19% (54) 95.21% (92) 95.99% (77) 95.99% (77) 96.41% (69) 90.42% (184)

c-transformed 99.55% (19) 99.12% (37) 99.12% (37) 99.12% (37) 99.17% (35) 96.97% (128)

All 98.23% 98.19% 98.72% 98.72% 98.82% 95.24%

(102) (104) (74) (74) (68) (274)

In this case the squared v2 minimum distance classifier gives in average the best results and then the SVM. If we consider all the cases with classes represented by multiple textures, we computed the average classification rate: the best results are obtained with SVM (with a total of 399 errors), then follows the squared v2 distance (460 errors), the Topsøe, Jensen difference distances make 472 errors each, Jeffreys yields 616 misidentifications and the cosine distance makes a total of 1352 errors. For the situation of class content with mixture of all types of images we computed for each texture the total number of misidentifications, the results are summarized in Table 10. It is surprising that a texture such as Brick which has good saliency, with Table 10. Number of misidentifications Rotations Bark 64 Brick 169 Grass 21 Leather 8 Pigskin 19 Raffia 13 Sand 121 Straw 30 Water 0 Weave 51 Wood 3 Wool 54

c-transformed 85 27 56 19 5 2 26 25 0 24 0 24

All 161 146 61 33 23 6 113 21 0 76 4 52

554

A. Ignat and M. Luca

very distinct visual and directional characteristics has one of the highest misidentification rate and the Water texture which visually is very similar with Sand or Pigskin has no errors in classification. We tested the cosine, correlation, Topsøe, Jensen difference and squared v2 histogram similarity measures upon the six real world textures from Figs. 2 and 4. There are three classes of textures (bark, grass, and pavement) each class having two samples. We computed for all these textures the similarity measures to the other five images. In Table 11 we marked ‘T’ (true) when the minimum distance was achieved for the image in the same class and ‘F’ (false) otherwise. The DTCWT -based similarity measures provided the correct result in all the cases (the DTCWT column in Table 11). Table 11. Comparisons for real world images – distances applied to the histograms of the DTCWT coefficients and of the images Cosine DTCWT Bark T Grass T Pavement T

Correlation Img DTCWT Img F T F T T T F T F

Topsøe DTCWT T T T

Jensen Img DTCWT F T T T T T

Squared v2 Img DTCWT Img F T F T T T T T T

We compare the classification obtained applying the five distances to the histogram of the DTCWT coefficients to the same classification performed when computing the similarity of the histogram of the textures (the Img column in Table 11). We obtained the correct texture when comparing the histogram of the rotated textures with the not rotated ones, in almost all cases, as expected. The cosine and the correlation distances had 1 misidentification each for the 30° case.

6 Conclusions In this paper, we studied texture identification in digital images with a new method. We employed the Dual Tree Complex Wavelet Transform (DTCWT) for texture feature extraction and identification using various similarity measures. The DTCWT provides a set of six complex selective directional coefficients, for which we employed their magnitudes {Sj,d; j = 1,…,6, d = 1,..,D}. We computed the histograms of these coefficients and classified 12 textures from the Brodatz album using SVM or minimum distance classifiers employing six similarity measures: cosine and Pearson correlation coefficient, Jeffreys, Topsøe, Jensen difference and squared v2. Our main focus was on the influence of rotation and c-transformation on the capacity of texture recognition. We analysed several cases: classes with only one texture or with multiple textures, the rotations and c-transformations were treated separately or together. We also tested the proposed method on real world images with good identification results. Our further research will concentrate on testing the influence of the image size, the noise and also on other types of image transformations (blur, shear, etc.) beside rotations and

Tips on Texture Evaluation with Dual Tree Complex Wavelet Transform

555

c-transformations on the capacity of the presented methods for texture identification. These procedures will be tested also on image segmentation and indexing based on texture recognition.

References 1. Petrou, M., Sevilla, P.G.: Image Processing: Dealing With Texture. Wiley, England (2006) 2. Mirmehdi, M., Xie, X., Suri, J. (eds.) Handbook of Texture Analysis. Imperial College Press, London (2008) 3. Wang, X., Georganas, N.D., Petriu, E.M.: Fabric texture analysis using computer vision techniques. IEEE Trans. Instrum. Measur. 60(1), 44–56 (2011) 4. Basile, T.M.A., Caponetti, L., Castellano, G., Sforza, G.: A texture-based image processing approach for the description of human oocyte cytoplasm. IEEE Trans. Instrum. Measur. 59 (10), 2591–2601 (2010) 5. Kingsbury, N.G.: The dual-tree complex wavelet transform: a new technique for shift invariance and directional filters. In: Proceedings of the 8th IEEE DSP Workshop, Paper no. 86, Utah, 9–12 August 1998 6. Kingsbury, N.G.: The dual-tree complex wavelet transform: a new efficient tool for image restoration and enhancement. In: Proceedings of the European Signal Processing Conference, Rhodes, September 1998, pp. 319–322 (1998) 7. Brodatz, P.: Textures: a photographic album for artists and designers, Dover, New York, USA (2014). http://www.ux.uis.no/*tranden/brodatz.html 8. Hill, P.R., Bull, D.R., Canagarajah, C.N.: Rotationally invariant texture features using the dual-tree complex wavelet transform. In: IEEE Proceedings of the International Conference on Image Processing (ICIP), Vancouver, Canada, vol. 3, pp. 901–904 (2000) 9. Selesnick, I.W., Baraniuk, R.G., Kingsbury, N.G.: The dual tree complex wavelet transform DTCWT. IEEE Sig. Process. Mag. 22(6), 123–151 (2005) 10. Kingsbury, N.G.: Complex wavelets for shift invariant analysis and filtering of signals. J. Appl. Comput. Harmonic Anal. 3, 234–253 (2001) 11. Kingsbury, N.G.: Design of Q-shift complex wavelets for image processing using frequency domain energy minimization. In: Proceedings of the IEEE Conference on Image Processing, Barcelona, 15–17 September (2003). Paper 1199 12. Mallat, S.G.: A theory for multiresolution signal decomposition: the wavelet representation. IEEE Trans. Pattern Anal. Mach. Intell. 11(7), 674–693 (1989) 13. Hatipoglu, S., Mitra, S.K., Kingsbury, N.G.: Image texture description using complex wavelet transform. In: Proceedings of the IEEE International Conference on Image Processing, Vancouver, BC, Canada, September 2000, vol. 2, pp. 530–533 (2000) 14. Wang, H.-Z., He, X.-H., Zai, W.-J.: Texture image retrieval using dual-tree complex wavelet transform. In: Proceedings of the 2007 International Conference on Wavelet Analysis and Pattern Recognition, Beijing, China, 2–4 November 2007, pp. 230–234 (2007) 15. Mumtaz, A., Gilani, M.S.A., Hameed, K., Jameel, T.: Enhancing performance of image retrieval systems using dual tree complex wavelet transform and support vector machines. J. Comput. Inf. Technol. (CIT) 16(1), 57–68 (2008) 16. Celik, T., Tjahadi, T.: Multiscale texture classification using dual-tree complex wavelet transform. Pattern Recogn. Lett. 30, 331–339 (2009) 17. Hatipoglu, S., Mitra, S.K., Kingsbury, N.: Texture classification using dual-tree complex wavelet transform. IEEE Image Process. Appl. 465, 344–347 (1999)

556

A. Ignat and M. Luca

18. Lo, E.H.S., Pickering, M., Frater, M., Arnold, J.: Scale and rotation invariant texture features from the dual-tree complex wavelet transform. In: IEEE Proceedings of the International Conference on Image Processing (ICIP), 24–27 October, Singapore, vol. 1, pp. 227–230 (2004) 19. Chen, S., Shang, Y., Mao, B., Lian, Q.: Rotation invariant texture classification algorithm based on DT-CWT and SVM. In: Liu, D., Fei, S., Hou, Z., Zhang, H., Sun, C. (eds.) Proceedings of the 4th International Symposium on Neural Networks Advances in Neural Networks ISNN 2007, pp. 454–460. Springer, Heidelberg (2007). 20. Cha, S.-H.: Comprehensive survey on distance/similarity measures between probability density functions. Int. J. Math. Models Methods Appl. Sci. 1(4), 300–307 (2007) 21. Xu, R., WunschII, D.C.: Clustering. IEEE Press/Wiley, Hoboken (2009) 22. Gonzales, R.C., Woods, R.E.: Digital Image Processing, 3rd ed. Chap. 3, pp. 132–137. Prentice Hall, Upper Saddle River (2008) 23. Costin (Luca), M., Ignat, A.: Pitfalls in using dual tree complex wavelet transform for texture featuring: a discussion. In: IEEE WISP 2011 - 7th IEEE International Symposium on Intelligent Signal Processing, Floriana, Malta, 19–21 September 2011, pp. 110–115 (2011) 24. Kingsbury’s, N.G.: http://www-sigproc.eng.cam.ac.uk/Main/NGK

Author Index

A Abdmouleh, Med Karim, 164, 524 Ahmed, Adeel, 216, 229, 253 Ahmed, Zaheer, 216, 229, 253 Aiordachioaie, Dorel, 430 Amri, Hedi, 164, 524 Arif, Muhammad, 216, 229, 253 Ashour, Amira S., 367, 375, 394 Azizi, Nabiha, 375 B Badkoobe, Marzieh, 189 Balas, Marius M., 12, 290 Balas, Sanda V., 367 Balas, Valentina Emilia, 78, 189, 316, 349 Banu, P.K. Nizar, 408 Barbu, Tudor, 423 Barbulescu, C., 47 Bărbuț, Alexandra, 308 Bărbuț, Liliana, 308 Beiu, Valeriu, 349 Bejinariu, Silviu-Ioan, 509 Ben Abdessalem Karaa, Wahiba, 394 Bold, Nicolae, 135 Boran, Robert A., 12 Bouhlel, Med Salim, 164, 497, 524 Bouhlel, Mohamed Salim, 474 C Chaki, Nabendu, 385 Cheriguene, Soraya, 375 Chiuchisan, Iulian, 339 Chiuchisan, Iuliana, 265, 327, 339, 349 Cornel, Barna, 385 Costin, Hariton, 509 Covasa, Mihai, 265 Crișan-Vida, Mihaela, 308

D David, Bertrand, 524 De Luise, D. López, 359 Dey, Nilanjan, 367, 375, 394 Dineva, Adrienn, 446, 455 Doloc, Cris, 265 Dombi, József, 446 Domșa, Ovidiu, 135 Domuschiev, I., 367 Dragan, Florin, 176 Dzitac, Simona, 60, 95, 107 E Elloumi, Nessrine, 497 Elnaggar, Ola E., 203 F Fayek, Magda B., 203 Filip, Ioan, 70, 176 G Gal, Andrei, 176 Gal-Nadasan, Emanuela, 298 Gal-Nadasan, Norbert, 298 Gargouri, Malek, 524 Geman, Oana, 265, 327, 339, 349 Glielmo, Luigi, 118 Gospodinov, M., 367 Gospodinova, E., 367 H Haghani, M., 316 Hemanth, D.J., 359 Holhjem, Øystein Hov, 118 Hosseinabadi, Ali Asghar Rahmani, 189

© Springer International Publishing AG 2018 V.E. Balas et al. (eds.), Soft Computing Applications, Advances in Intelligent Systems and Computing 633, DOI 10.1007/978-3-319-62521-8

557

558

Author Index

I Iannelli, Luigi, 118 Ignat, Anca, 540 Indumathi, G., 107 Isoc, Dorin, 18, 34

O Olariu, Iustin, 394, 408 Olariu, Teodora, 367, 375, 408 Own, Hala S., 408

J Jabbar, Muhammad, 216, 229, 253 Jaberi, O., 316 Jurcău, Daniel-Alexandru, 276

P Palmieri, Giovanni, 118 Park, J.S., 359 Parthasarathy, S., 60 Paul, Swagata, 385 Pérez, J., 359 Piuri, Vincenzo, 455 Poenaru, Dan V., 298 Popa-Andrei, Diana, 298 Popescu, Doru Anastasiu, 135 Prabhat, Purnendu, 487 Prostean, Octavian, 70

K Khajuria, Garvita, 487 Khalfallah, Ali, 164, 524 Kher, Vansha, 487 Kilyeni, St., 47 Kohli, Meena, 487 Kotta, Ülle, 78 Kumar, Ravish, 107 L Lazăr, Camelia, 509 Lolea, Marius, 95 Loukil Hadj Kacem, Habiba, 497 Loukil, Habiba, 474 Luca, Mihaela, 540 Luca, Ramona, 509 M Madan, Vinod K., 487 Maffei, Alessio, 118 Mannai, Monia, 394 Marafioti, Giancarlo, 118 Mathisen, Geir, 118 Miclea, Razvan-Catalin, 3 Milici, Laurentiu-Dan, 265, 327, 339 Milici, Mariana-Rodica, 265, 327 Milici, Rodica-Mariana, 339 Mnerie, Corina A., 375 Mohanna, Farahnaz, 189 Mortazavi, S.A.R., 316 Mortazavi, S.M.J., 316 Muniyasamy, K., 60 N Niţă, Cristina Diana, 509

R Rahali, Mourad, 474 Ramadan, Rabie A., 203 Ramasamy, Srini, 78 Rizwan, Abdul Rasheed, 216, 253 Rohatinovici, Noemi, 385 Rostami, Ali Shokouhi, 189 Rotaru, Florin, 509 S Sandru, Florin, 3 Saravanakumar, G., 78 Sarkar, Bidyut Biman, 385 Shi, Fuqian, 375 Silea, Ioan, 3 Simo, A., 47 Sofrag, Adelin, 290 Soleimani, A., 316 Srinivasan, Seshadhri, 60, 78, 107, 118 Srivastava, Juhi R., 143 Stoicu-Tivadar, Lăcrămioara, 308 Stoicu-Tivadar, Vasile, 276, 298 Subathra, B., 60, 78 Sudarshan, T.S.B., 143 Sundarajan, N., 78 Szeidert, Iosif, 70

Author Index

559

T Tar, József K., 455 Toderean (Aldea), Roxana, 349 Tripathi, Shikha, 462

V Várkonyi-Kóczy, Annamária R., 455 Vasar, Cristian, 70 Vekkot, Susmitha, 462

U Ullah, Muhammad Sami, 229, 253

Z Zamani, A., 316